[ { "id": "3d-ai-generation", "name": "3D AI Generation", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "multimodal" }, { "to": "digital-human" }, { "to": "computer-vision" } ], "track": "today", "i18n": { "en": { "fullName": "3D AI Generation", "factExplain": "AI that creates 3D models or scenes automatically.", "humanExplain": "It is like a magic pottery wheel on your desk. Say “tiny treehouse,” and it spins up a rough 3D version.\n\nIt makes first drafts for games, art, and product design. People still fix the small details by hand.", "humanExplainDisplay": "It is like a ==magic pottery wheel== on your desk.\nSay “tiny treehouse,”\nand it spins up a ==rough 3D version==.\n\nIt makes first drafts for games, art,\nand product design.\nPeople still fix the small details by hand.", "relationsNarrative": "Diffusion\nMany 3D generation methods use Diffusion to make shapes or textures.\n\nMultimodal AI\nIt often works with text, images, and 3D information together.\n\nDigital human\n3D generation can provide character looks and scene pieces for a digital human.\n\nComputer Vision\nIt often uses Computer Vision to understand structure and space in images.", "relations": { "diffusion": { "label": "often uses …", "note": "Many 3D tools use Diffusion to create shapes or textures." }, "multimodal": { "label": "is a use of …", "note": "It often mixes text, images, and 3D information." }, "digital-human": { "label": "helps build …", "note": "It can make character looks and scene pieces faster." }, "computer-vision": { "label": "leans on …", "note": "Computer Vision helps it understand object shape and space." } } }, "zh": { "fullName": "3D AI 生成", "factExplain": "用 AI 自动生成三维模型或场景的技术。", "humanExplain": "以前你还得自己拼乐高说明书，现在 AI 像班里那个立体手工王，听一句就把桌椅房间先搭成型。\n\n常用于游戏、美术和产品设计，加速建模起稿，细节精修仍常要人工。", "humanExplainDisplay": "以前你还得自己拼乐高说明书，\n现在 AI 像班里那个\n==立体手工王==，\n听一句就把桌椅房间先==搭成型==。\n\n常用于游戏、美术和产品设计，\n加速建模起稿，细节精修仍常要人工。", "relationsNarrative": "Diffusion\n不少 3D 生成方法会借扩散模型生成形状或纹理。\n\nMultimodal\n它常结合文字、图片与三维信息一起工作。\n\nDigital Human\n3D 生成能为数字人提供角色外形和场景资产。\n\nComputer Vision\n它常借助视觉能力理解图片里的结构与空间。", "relations": { "diffusion": { "label": "常借…生成", "note": "不少 3D 方案会借扩散模型出形状。" }, "multimodal": { "label": "属于…应用", "note": "它常同时处理文字、图片与三维信息。" }, "digital-human": { "label": "可用于打造…", "note": "它能更快生成角色外形与场景资产。" }, "computer-vision": { "label": "依赖…理解", "note": "它常靠视觉能力还原物体结构。" } } } } }, { "id": "a-search", "name": "A* Search", "layer": "L2", "era": "1968", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "graph-search" }, { "to": "heuristic-search" }, { "to": "dynamic-programming" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "A* Search Algorithm", "factExplain": "A* finds the best path using past cost and a smart distance guess.", "humanExplain": "A* is like a kid in a corn maze. It counts muddy steps, then follows the path smelling most like popcorn.\n\nYou meet it in game maps and robot routes. It finds a solid path fast and skips many dead ends.", "humanExplainDisplay": "A* is like a kid in a ==corn maze==.\nIt counts ==muddy steps==,\nthen follows the path smelling most like popcorn.\n\nYou meet it in game maps and robot routes.\nIt finds a solid path fast\nand skips many dead ends.", "relationsNarrative": "Graph Search\nA* is a classic shortest-path method in Graph Search.\n\nHeuristic Search\nA* uses a heuristic guess to try the best-looking path first.\n\nDP\nA* keeps the cost so far, so it feels a bit like DP.\n\nAgent\nAn Agent can use A* to find a route or solve a task.", "relations": { "graph-search": { "label": "belongs to …", "note": "A* is a classic method in Graph Search." }, "heuristic-search": { "label": "uses … guesses", "note": "The heuristic guess helps A* choose promising paths first." }, "dynamic-programming": { "label": "borrows from …", "note": "A* keeps the cost so far before it chooses again." }, "agent": { "label": "can guide an …", "note": "An Agent can use A* to plan a route or a task." } } }, "zh": { "fullName": "A* 搜索算法", "factExplain": "一种结合已走代价与启发估计的最优搜索算法。", "humanExplain": "A*找路像外卖骑手赶单：既算已经绕了多少街口，也掂量前面哪条巷子更近，不会瞎冲。\n\n它常用于寻路、路径规划和任务求解，能更快找到靠谱路线。", "humanExplainDisplay": "A*找路像外卖骑手赶单：\n既算已经==绕了多少街口==，\n也掂量前面==哪条巷子更近==，\n不会瞎冲。\n\n它常用于寻路、路径规划\n和任务求解，\n能更快找到靠谱路线。", "relationsNarrative": "Graph Search\nA* Search 是图搜索中的经典最短路方法。\n\nHeuristic Search\n它靠启发式估计优先探索更有希望的路径。\n\nDynamic Programming\n它会结合累计代价，带有动态规划味道。\n\nAgent\nAgent 做路径规划或任务规划时可用它找解。", "relations": { "graph-search": { "label": "属于…方法", "note": "它是图搜索里的经典代表。" }, "heuristic-search": { "label": "依赖…估路", "note": "启发信息决定它搜得快不快。" }, "dynamic-programming": { "label": "借鉴…思想", "note": "它会累计历史代价再做选择。" }, "agent": { "label": "可为…找路", "note": "Agent 做规划时常可用它。" } } } } }, { "id": "a3c", "name": "A3C", "layer": "L2", "era": "2016", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "actor-critic" }, { "to": "deep-reinforcement-learning" }, { "to": "policy-gradient" }, { "to": "on-policy-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Asynchronous Advantage Actor-Critic", "factExplain": "A Deep RL method where many workers learn in parallel without waiting.", "humanExplain": "A3C is like a video game speedrun club. Several players try routes at once. Then they shout tips across the couch.\n\nIt helped train game AIs faster. It also helped control things inside simulators.", "humanExplainDisplay": "A3C is like a ==video game speedrun club==.\nSeveral players try routes at once.\nThen they ==shout tips across the couch==.\n\nIt helped train game AIs faster.\nIt also helped control things inside simulators.", "relationsNarrative": "Actor-Critic\nA3C is the async parallel version of Actor-Critic.\n\nDeep RL\nA3C was an early landmark algorithm in Deep RL.\n\nPolicy Gradient\nA3C uses Policy Gradient to update the actor's action policy.\n\nOn-Policy Learning\nA3C usually samples with the current policy and learns right away.", "relations": { "actor-critic": { "label": "adds async workers to …", "note": "A3C trains many Actor-Critic workers in parallel." }, "deep-reinforcement-learning": { "label": "is a classic …", "note": "A3C was an early landmark Deep RL algorithm." }, "policy-gradient": { "label": "updates policy with …", "note": "Policy Gradient improves the actor's action choices." }, "on-policy-learning": { "label": "belongs to …", "note": "A3C mainly learns from fresh runs by the current policy." } } }, "zh": { "fullName": "异步优势演员-评论家（Asynchronous Advantage Actor-Critic）", "factExplain": "一种异步并行的深度强化学习算法。", "humanExplain": "A3C 像麻将馆复盘：几桌同时试牌路，谁胡谁放炮，立刻喊给全场听。\n\n曾加速游戏智能体训练，也用于仿真控制。", "humanExplainDisplay": "A3C 像麻将馆复盘：\n几桌同时==试牌路==，\n谁胡谁放炮，\n立刻==喊给全场听==。\n\n曾加速游戏智能体训练，\n也用于仿真控制。", "relationsNarrative": "Actor-Critic\nA3C 是 Actor-Critic 的异步并行版本。\n\nDeep RL\nA3C 是深度强化学习早期的代表算法。\n\nPolicy Gradient\nA3C 用策略梯度更新演员的动作策略。\n\nOn-Policy Learning\nA3C 通常用当前策略采样并立即学习。", "relations": { "actor-critic": { "label": "异步扩展…", "note": "A3C 把多个演员并行训练。" }, "deep-reinforcement-learning": { "label": "代表…", "note": "它是早期深度强化学习标志算法。" }, "policy-gradient": { "label": "用…更新策略", "note": "策略梯度负责改进动作选择。" }, "on-policy-learning": { "label": "属于…", "note": "它主要用当前策略采样学习。" } } } } }, { "id": "active-learning", "name": "Active Learning", "layer": "L2", "era": "1994", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "data-labeling" }, { "to": "model-uncertainty" }, { "to": "human-in-the-loop" } ], "track": "history", "i18n": { "en": { "fullName": "Active Learning", "factExplain": "A way to train models by labeling only the most useful examples.", "humanExplain": "Active Learning is like a kid in math class. They skip 2+2 and ask the teacher about the monster problem.\n\nIt helps when labels cost money. The model asks people about fewer examples, but learns faster.", "humanExplainDisplay": "Active Learning is like a ==kid in math class==.\nThey skip ==2+2==\nand ask the teacher\nabout the monster problem.\n\nIt helps when labels cost money.\nThe model asks people about fewer examples,\nbut learns faster.", "relationsNarrative": "Supervised Learning\nActive Learning picks more useful examples for Supervised Learning.\n\nData Labeling\nActive Learning spends the labeling budget on examples worth asking about.\n\nModel uncertainty\nThe model asks first about examples it is unsure about.\n\nHuman-in-the-loop\nActive Learning needs people to label the key examples.", "relations": { "supervised-learning": { "label": "picks examples for …", "note": "Active Learning usually chooses examples for supervised training." }, "data-labeling": { "label": "cuts … costs", "note": "It spends human labeling time on the examples worth asking about." }, "model-uncertainty": { "label": "uses … to choose", "note": "Uncertain examples often get sent to humans first." }, "human-in-the-loop": { "label": "depends on … feedback", "note": "People give the key labels the model needs." } } }, "zh": { "fullName": "主动学习", "factExplain": "让模型挑最值得标注样本的训练方法。", "humanExplain": "主动学习像老师划重点：不刷整本练习册，只拎最懵的题问人。\n\n标注贵时最有用，少请人也能把模型教准。", "humanExplainDisplay": "主动学习像老师划重点：\n不刷整本练习册，\n只拎==最懵的题==问人。\n\n标注贵时最有用，\n少请人，\n也能把模型教准。", "relationsNarrative": "Supervised Learning\n主动学习通常为监督训练挑选更有价值的样本。\n\nData Labeling\n它把标注预算集中在最值得问的样本上。\n\nModel Uncertainty\n模型越拿不准的样本，越可能被优先标注。\n\nHuman-in-the-loop\n主动学习需要人类为关键样本提供标签。", "relations": { "supervised-learning": { "label": "服务于…", "note": "主动学习通常为监督训练挑样本。" }, "data-labeling": { "label": "减少…成本", "note": "它把人工标注用在刀刃上。" }, "model-uncertainty": { "label": "依据…挑样本", "note": "不确定性常用来判断该问谁。" }, "human-in-the-loop": { "label": "依赖…反馈", "note": "人类标注者提供关键答案。" } } } } }, { "id": "actor-critic", "name": "Actor-Critic", "layer": "L2", "era": "1983", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "policy-gradient" }, { "to": "q-learning" }, { "to": "reinforcement-learning" }, { "to": "rlhf" } ], "track": "history", "i18n": { "en": { "fullName": "Actor-Critic", "factExplain": "A reinforcement learning method with one part choosing actions and one part scoring them.", "humanExplain": "One kid tries dance moves. The judge holds a scorecard: “Less robot arms.”\n\nIn AI, the actor picks the move. The critic scores it, so robots, games, and RLHF train with less random guessing.", "humanExplainDisplay": "One kid tries dance moves.\nThe judge holds a ==scorecard==:\n“==Less robot arms==.”\n\nIn AI, the actor picks the move.\nThe critic scores it,\nso robots, games, and RLHF train\nwith less random guessing.", "relationsNarrative": "Policy Gradient\nActor-Critic adds a value score to Policy Gradient, so training is steadier.\n\nQ-Learning\nIt borrows value learning, but it also learns the actions directly.\n\nRL\nActor-Critic is a classic RL method family, often used for control.\n\nRLHF\nMany RLHF systems use it to improve the policy and keep learning stable.", "relations": { "policy-gradient": { "label": "stabilizes … training", "note": "It adds a value score, so Policy Gradient jumps around less." }, "q-learning": { "label": "borrows value learning from …", "note": "It learns a policy and a value score at the same time." }, "reinforcement-learning": { "label": "is a classic … method", "note": "It is a classic method family in RL." }, "rlhf": { "label": "helps optimize …", "note": "It helps RLHF update the policy without swinging too wildly." } } }, "zh": { "fullName": "演员-评论家方法", "factExplain": "一种把选动作与打分估值分开的强化学习方法。", "humanExplain": "跟说相声似的：一个负责现挂出活，一个负责台下听响，别让演员自己给自己鼓掌。\n\n常用于机器人、游戏和 RLHF，让动作学习更稳、更少乱试。", "humanExplainDisplay": "跟说相声似的：\n一个负责==现挂出活==，\n一个负责台下听响，\n别让演员==自己给自己鼓掌==。\n\n常用于机器人、\n游戏和 RLHF，\n让动作学习更稳、更少乱试。", "relationsNarrative": "Policy Gradient\n它在策略梯度上加入价值评估，训练通常更稳。\n\nQ-learning\n它吸收价值学习思路，但不只学打分，也直接学动作。\n\nReinforcement-learning\n它是强化学习中的经典方法族，常用于连续控制。\n\nRLHF\n很多对齐训练会用它来优化策略并稳定学习过程。", "relations": { "policy-gradient": { "label": "改进…训练", "note": "在策略梯度上加估值辅助降方差。" }, "q-learning": { "label": "结合…思路", "note": "一边学策略，一边学价值评估。" }, "reinforcement-learning": { "label": "属于…方法", "note": "它是强化学习里的经典路线。" }, "rlhf": { "label": "支撑…优化", "note": "常被用来稳定偏好对齐训练。" } } } } }, { "id": "adaboost", "name": "AdaBoost", "layer": "L2", "era": "1995", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "ensemble-learning" }, { "to": "decision-tree" }, { "to": "gradient-boosting" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Adaptive Boosting", "factExplain": "An ensemble method with repeated reweighting of misclassified examples.", "humanExplain": "AdaBoost is like a coach at free-throw practice. Miss the shot, and the next ball comes straight back to you.\n\nIt trains simple models again and again, so past misses get extra weight. It often sorts things into classes, but noisy data can make it chase junk.", "humanExplainDisplay": "AdaBoost is like a coach at free-throw practice.\nMiss the shot,\nand ==the next ball== comes ==straight back to you==.\n\nIt trains simple models again and again,\nso past misses get extra weight.\nIt often sorts things into classes,\nbut noisy data can make it chase junk.", "relationsNarrative": "Ensemble\nAdaBoost combines many weak learners into one stronger predictor.\n\nDecision Tree\nAdaBoost often trains shallow Decision Trees as weak learners.\n\nGradient Boosting\nGradient Boosting follows the same idea of fixing mistakes round by round.\n\nClassification\nAdaBoost is often used to put examples into the right class.", "relations": { "ensemble-learning": { "label": "belongs to …", "note": "AdaBoost is a classic boosting ensemble method." }, "decision-tree": { "label": "often uses … as weak learners", "note": "Shallow trees are often combined again and again." }, "gradient-boosting": { "label": "inspired …", "note": "Both fix earlier mistakes one round at a time." }, "classification": { "label": "often used for …", "note": "AdaBoost is most often used for classification tasks." } } }, "zh": { "fullName": "自适应提升", "factExplain": "一种反复重权误分类样本的集成学习算法。", "humanExplain": "AdaBoost像班主任盯错题：谁总丢分就加练，几个普通同学也能组学霸队。\n\n常做分类；小模型变强，噪声多会跑偏。", "humanExplainDisplay": "AdaBoost像班主任==盯错题==：\n谁总丢分就加练，\n几个普通同学\n也能组==学霸队==。\n\n常做分类；\n小模型变强，\n噪声多会跑偏。", "relationsNarrative": "Ensemble\n它通过组合多个弱学习器，得到更强预测器。\n\nDecision Tree\n浅层决策树常被它当作弱学习器反复训练。\n\nGradient Boosting\n梯度提升继承了逐轮改错的提升思想。\n\nClassification\n它最常用于把样本分到正确类别。", "relations": { "ensemble-learning": { "label": "属于…", "note": "它是经典的提升式集成方法。" }, "decision-tree": { "label": "常用…做弱学习器", "note": "浅层树常被它反复组合。" }, "gradient-boosting": { "label": "启发…", "note": "二者都逐轮修正前一轮错误。" }, "classification": { "label": "常用于…", "note": "它最常见场景是分类任务。" } } } } }, { "id": "adagrad", "name": "AdaGrad", "layer": "L2", "era": "2011", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "gradient-descent" }, { "to": "sgd" }, { "to": "rmsprop" }, { "to": "adam" } ], "track": "history", "i18n": { "en": { "fullName": "Adaptive Gradient Algorithm", "factExplain": "An optimizer sets each parameter's learning rate from past gradients.", "humanExplain": "AdaGrad is like a gym coach with a tiny clipboard. Busy muscles get lighter weights. Lazy muscles get the spotlight.\n\nIt helps with rare clues, like rare words. But its steps keep shrinking. Learning can crawl near the end.", "humanExplainDisplay": "AdaGrad is like a ==gym coach==\nwith a tiny clipboard.\n==Busy muscles== get lighter weights.\nLazy muscles get the spotlight.\n\nIt helps with rare clues,\nlike rare words.\nBut its steps keep shrinking.\nLearning can crawl near the end.", "relationsNarrative": "Gradient Descent\nAdaGrad follows the same gradient direction, but each parameter uses its own step size.\n\nSGD\nSGD often uses one shared step size, but AdaGrad changes it per parameter.\n\nRMSProp\nRMSProp uses a moving average to slow AdaGrad's shrinking step sizes.\n\nAdam\nAdam keeps AdaGrad's adaptive step sizes and adds momentum estimates.", "relations": { "gradient-descent": { "label": "refines … step size", "note": "AdaGrad still moves along the gradient direction." }, "sgd": { "label": "changes … learning rate", "note": "AdaGrad changes one shared step size into one size per parameter." }, "rmsprop": { "label": "inspires … fix", "note": "RMSProp uses a moving average so steps shrink less fast." }, "adam": { "label": "inspires … adaptive steps", "note": "Adam keeps custom step sizes and adds momentum." } } }, "zh": { "fullName": "自适应梯度算法", "factExplain": "按历史梯度为每个参数自调学习率的优化算法。", "humanExplain": "AdaGrad像课堂点名：常举手的少叫，沉默同学多给镜头。\n\n适合稀疏特征训练；但步长会衰减，后期易学慢。", "humanExplainDisplay": "AdaGrad像课堂点名：\n==常举手的少叫==，\n沉默同学==多给镜头==。\n\n适合稀疏特征训练；\n但步长会衰减，\n后期易学慢。", "relationsNarrative": "Gradient Descent\nAdaGrad 仍沿梯度方向走，只是每个参数步长不同。\n\nSGD\nSGD 常用统一步长，AdaGrad 改成逐参数调节。\n\nRMSProp\nRMSProp 用滑动平均缓解 AdaGrad 步长衰减过快。\n\nAdam\nAdam 继承自适应步长思想，并加入动量估计。", "relations": { "gradient-descent": { "label": "细化…的步长", "note": "仍沿梯度方向更新参数。" }, "sgd": { "label": "改造…的学习率", "note": "把统一步长改成逐参数调节。" }, "rmsprop": { "label": "启发…的改进", "note": "滑动平均缓解步长过快变小。" }, "adam": { "label": "启发…的自适应", "note": "Adam 继续融合自适应步长。" } } } } }, { "id": "adaline", "name": "ADALINE", "layer": "L3", "era": "1960", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "widrow-hoff-learning-rule" }, { "to": "perceptron" }, { "to": "gradient-descent" }, { "to": "neural-network" } ], "track": "history", "i18n": { "en": { "fullName": "Adaptive Linear Neuron", "factExplain": "A one-layer network trained by reducing error in its linear output.", "humanExplain": "ADALINE is like a picky DJ with a volume knob. If the music is three clicks too loud, it turns down three clicks.\n\nIt shows up in simple filters and classifiers. It keeps changing its weights to shrink the error.", "humanExplainDisplay": "ADALINE is like a ==picky DJ==\nwith a ==volume knob==.\nIf the music is three clicks too loud,\nit turns down three clicks.\n\nIt shows up in simple filters and classifiers.\nIt keeps changing its weights\nto shrink the error.", "relationsNarrative": "Widrow-Hoff Learning Rule (Least Mean Squares Rule)\nADALINE is classically trained with the Widrow-Hoff Learning Rule.\n\nPerceptron\nADALINE looks like a Perceptron, but it learns from the size of the error.\n\nGradient Descent\nADALINE updates its weights like repeated steps of Gradient Descent.\n\nNeural-network\nADALINE is a classic early one-layer Neural-network.", "relations": { "widrow-hoff-learning-rule": { "label": "is trained by …", "note": "Its classic training rule is LMS." }, "perceptron": { "label": "learns differently from …", "note": "ADALINE uses the size of the error, not just right or wrong." }, "gradient-descent": { "label": "can be understood as …", "note": "Its weights move toward lower error." }, "neural-network": { "label": "is an early …", "note": "ADALINE is a classic one-layer linear network." } } }, "zh": { "fullName": "Adaptive Linear Neuron，自适应线性神经元", "factExplain": "一种用线性输出和误差最小化训练的单层网络。", "humanExplain": "ADALINE 像煎饼摊调酱：不只听“咸了”，还按差几勺慢慢补到刚好。\n\n用于自适应滤波和分类，按误差持续校正权重。", "humanExplainDisplay": "ADALINE 像\n==煎饼摊调酱==：\n不只听“咸了”，\n还按==差几勺==慢慢补到刚好。\n\n用于自适应滤波和分类，\n按误差持续校正权重。", "relationsNarrative": "Widrow-Hoff Learning Rule\nADALINE 的经典训练法就是 Widrow-Hoff 规则。\n\nPerceptron\n它外形像感知机，但按连续误差学习。\n\nGradient Descent\n它的权重更新可看作一次次梯度下降。\n\nNeural Network\n它是早期单层神经网络的代表。", "relations": { "widrow-hoff-learning-rule": { "label": "用…训练", "note": "经典训练法就是 LMS。" }, "perceptron": { "label": "区别于…", "note": "不只看对错，还看误差。" }, "gradient-descent": { "label": "可由…理解", "note": "权重沿误差变小方向调。" }, "neural-network": { "label": "属于早期…", "note": "单层线性网络的代表。" } } } } }, { "id": "adam", "name": "Adam", "layer": "L2", "era": "2014", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "parameter" }, { "to": "pretraining" }, { "to": "fine-tuning" } ], "track": "history", "i18n": { "en": { "fullName": "Adaptive Moment Estimation optimizer", "factExplain": "A neural network optimizer that automatically changes its learning steps.", "humanExplain": "Adam is like smart sneakers at school. They tiptoe on slick floors, then zoom on carpet.\n\nIt updates model parameters and keeps training steadier. You often meet it in pretraining and fine-tuning.", "humanExplainDisplay": "Adam is like ==smart sneakers== at school.\nThey ==tiptoe on slick floors==,\nthen zoom on carpet.\n\nIt updates model parameters\nand keeps training steadier.\nYou often meet it in pretraining\nand fine-tuning.", "relationsNarrative": "Neural-network\nAdam updates weights so the neural network learns the task.\n\nParameter\nParameters are the parts Adam directly changes.\n\nPretraining\nPretraining often uses optimizers like Adam to update parameters.\n\nFine-tuning\nFine-tuning also uses Adam to adapt the model to a new task.", "relations": { "neural-network": { "label": "updates …", "note": "It changes weights so the neural network learns better." }, "parameter": { "label": "tunes …", "note": "Parameters are the things Adam changes each step." }, "pretraining": { "label": "often used in …", "note": "Large model pretraining often uses Adam to keep learning moving." }, "fine-tuning": { "label": "also used in …", "note": "Fine-tuning often uses Adam to adjust parameters for a new task." } } }, "zh": { "fullName": "自适应矩估计优化器", "factExplain": "一种会自动调学习步子的神经网络优化算法。", "humanExplain": "训练时它特别像会看火候的厨子：这边太猛就关小点，那边没熟就多加把劲。\n\n它用于更新模型参数，让训练更稳，常见于预训练和微调。", "humanExplainDisplay": "训练时它特别像\n会看火候的==厨子==：\n这边太猛就关小点，\n那边没熟就==多加把劲==。\n\n它用于更新模型参数，\n让训练更稳，\n常见于预训练和微调。", "relationsNarrative": "Neural-network\n它通过更新权重，推动神经网络逐步学会任务。\n\nParameter\n参数是它直接调整的对象，决定模型怎么变。\n\nPretraining\n预训练通常要靠它这类优化器持续更新参数。\n\nFine-tuning\n微调阶段也会继续用它来适配新任务。", "relations": { "neural-network": { "label": "负责更新…", "note": "它靠调整权重让神经网络越学越好。" }, "parameter": { "label": "专门调整…", "note": "参数就是它每一步要改的对象。" }, "pretraining": { "label": "常用于…训练", "note": "大模型预训练时常靠它推进学习。" }, "fine-tuning": { "label": "也用于…微调", "note": "微调阶段也常用它继续改参数。" } } } } }, { "id": "adamw", "name": "AdamW", "layer": "L2", "era": "2017", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "adam" }, { "to": "weight-decay" }, { "to": "optimization" }, { "to": "fine-tuning" } ], "track": "history", "seo": { "en": { "title": "What Is AdamW? Why This Optimizer Keeps Two Separate Books", "description": "AdamW separates weight decay from Adam's update step — like two ledgers that never mix. A plain-English explainer of why training gets steadier and overfitting drops." }, "zh": { "title": "AdamW 是什么?和 Adam 差在哪,一文看懂 — AI Rookies", "description": "把权重衰减从 Adam 更新里拆出来,像记两本互不串账的账本:训练更稳、过拟合更少。人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "AdamW Optimizer", "factExplain": "An optimizer that separates weight decay from Adam’s normal update step.", "humanExplain": "AdamW is like a coach with two scoreboards. Practice points go on one. Donut penalties go on another.\n\nIt is used to train and fine-tune models. It helps learning stay steady and helps models memorize less.", "humanExplainDisplay": "AdamW is like a coach with ==two scoreboards==.\nPractice points go on one.\n==Donut penalties== go on another.\n\nIt is used to train and fine-tune models.\nIt helps learning stay steady\nand helps models memorize less.", "relationsNarrative": "Adam\nAdamW keeps Adam’s adaptive updates, but fixes how weight decay is applied.\n\nWeight Decay\nAdamW separates Weight Decay from the normal gradient update.\n\nOptimization\nAdamW is a common optimizer for training neural networks.\n\nFine-tuning\nAdamW is often used as the default optimizer for fine-tuning large models.", "relations": { "adam": { "label": "improves … weight decay", "note": "It keeps Adam’s adaptive updates." }, "weight-decay": { "label": "decouples …", "note": "Weight Decay is applied to the parameters on its own." }, "optimization": { "label": "is an … method", "note": "It is a common optimizer for training neural networks." }, "fine-tuning": { "label": "often used for …", "note": "It is often the default optimizer for fine-tuning large models." } } }, "zh": { "fullName": "解耦权重衰减的 Adam 优化器", "factExplain": "一种将权重衰减从 Adam 更新中解耦的优化器。", "humanExplain": "AdamW 像记两本独立账本：一本记学到的本事、一本记该减的体重，各算各的、互不串账。\n\n常用于训练和微调，让收敛更稳、过拟合更少。", "humanExplainDisplay": "AdamW 像记\n==两本独立账本==：\n一本记学到的本事、\n一本记==该减的体重==，各算各的。\n\n常用于训练和微调，\n让收敛更稳、\n过拟合更少。", "relationsNarrative": "Adam\nAdamW 保留 Adam 的自适应更新，但修正权重衰减方式。\n\nWeight Decay\nAdamW 将权重衰减从梯度更新中单独拆出来。\n\nOptimization\nAdamW 是训练神经网络时常用的优化器之一。\n\nFine-tuning\n微调大模型时，AdamW 常被作为默认优化器。", "relations": { "adam": { "label": "改进…的减重方式", "note": "它保留 Adam 的自适应更新。" }, "weight-decay": { "label": "解耦…", "note": "权重衰减被单独作用于参数。" }, "optimization": { "label": "属于…方法", "note": "它是训练神经网络的常用优化器。" }, "fine-tuning": { "label": "常用于…", "note": "微调大模型时经常默认选它。" } } } } }, { "id": "adversarial-example", "name": "Adversarial Example", "layer": "L2", "era": "2013", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "ai-sandbox" }, { "to": "alignment" }, { "to": "prompt-injection" } ], "track": "history", "i18n": { "en": { "fullName": "Adversarial Example", "factExplain": "A tiny planned input change made to trick a model into a wrong answer.", "humanExplain": "It is like a tiny smudge on a barcode. You see cereal, but the scanner sees a lawn mower.\n\nPeople use adversarial examples in safety tests. They show how easily recognition models can be fooled.", "humanExplainDisplay": "It is like a ==tiny smudge== on a barcode.\nYou see cereal,\nbut the scanner sees a ==lawn mower==.\n\nPeople use adversarial examples in safety tests.\nThey show how easily recognition models can be fooled.", "relationsNarrative": "Computer Vision\nAdversarial examples are a classic way to fool image recognition models.\n\nAI sandbox\nAn AI sandbox lets people test adversarial examples in a safe space.\n\nAlignment\nAdversarial examples show weak spots in model stability and safety alignment.\n\nPrompt injection\nBoth try to trick a model, but one changes input and one changes instructions.", "relations": { "computer-vision": { "label": "often targets …", "note": "Image models were the first classic target for adversarial examples." }, "ai-sandbox": { "label": "can be tested in …", "note": "A sandbox lets people test the attack safely." }, "alignment": { "label": "exposes limits of …", "note": "Adversarial examples reveal weak spots in model safety and stability." }, "prompt-injection": { "label": "shares trickery with …", "note": "One tricks the input. The other tricks the instructions." } } }, "zh": { "fullName": "Adversarial Example（对抗样本）", "factExplain": "能诱导模型出错的刻意微小扰动输入。", "humanExplain": "它像外卖单上偷偷改一笔字，人看着还是那份饭，系统却可能把地址菜品全认岔。\n\n常用于安全测试，检查识别模型有多脆弱，也可能被用来攻击。", "humanExplainDisplay": "它像外卖单上\n偷偷改==一笔字==，\n人看着还是那份饭，\n系统却可能把地址菜品\n全认==岔==。\n\n常用于安全测试，\n检查识别模型有多脆弱，\n也可能被用来攻击。", "relationsNarrative": "Computer Vision\n对抗样本最经典的场景，就是误导图像识别模型。\n\nAI Sandbox\n它常在隔离环境里测试，用来观察模型会怎么被绕晕。\n\nAlignment\n它暴露模型在稳健性和安全对齐上的薄弱边界。\n\nPrompt Injection\n两者都在诱骗模型，只是一个改输入，一个改指令。", "relations": { "computer-vision": { "label": "常拿…开刀", "note": "图像模型最早最典型受其影响。" }, "ai-sandbox": { "label": "可在…里测试", "note": "适合隔离环境中演练攻击效果。" }, "alignment": { "label": "暴露…边界", "note": "说明模型稳健性与安全性不足。" }, "prompt-injection": { "label": "同属诱骗攻击", "note": "一个骗输入，一个骗指令。" } } } } }, { "id": "adversarial-training", "name": "Adversarial Training", "layer": "L2", "era": "2014", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "adversarial-example" }, { "to": "data-augmentation" }, { "to": "regularization" } ], "track": "history", "i18n": { "en": { "fullName": "Adversarial Training", "factExplain": "A way to train AI with trick examples, so it becomes harder to fool.", "humanExplain": "Adversarial Training is like dodgeball practice with one sneaky kid. They only throw the weird bounces, so game day feels less scary.\n\nYou see it in image recognition and security tools. It helps AI resist attacks, but training takes more work.", "humanExplainDisplay": "Adversarial Training is like ==dodgeball practice==\nwith ==one sneaky kid==.\nThey only throw the weird bounces,\nso game day feels less scary.\n\nYou see it in image recognition\nand security tools.\nIt helps AI resist attacks,\nbut training takes more work.", "relationsNarrative": "Adversarial Example\nAdversarial Training uses Adversarial Examples as practice for tricky inputs.\n\nData Augmentation\nAdversarial Training is a harsher form of Data Augmentation.\n\nRegularization\nAdversarial Training acts like Regularization by blocking weak shortcuts.", "relations": { "adversarial-example": { "label": "practices with …", "note": "Adversarial Examples are its main practice material." }, "data-augmentation": { "label": "is a harsher kind of …", "note": "It turns attack-style changes into training data." }, "regularization": { "label": "adds robustness like …", "note": "It uses harder examples to break weak shortcuts." } } }, "zh": { "fullName": "对抗训练", "factExplain": "把对抗样本加入训练以提升模型鲁棒性的方法。", "humanExplain": "对抗训练像武馆陪练专出阴招：先让模型挨几下刁钻招，真上擂台才不慌。\n\n用于图像识别和安全场景，提升抗攻击能力但更耗训练。", "humanExplainDisplay": "对抗训练像==武馆陪练专出阴招==：\n先让模型挨几下==刁钻招==，\n真上擂台才不慌。\n\n用于图像识别和安全场景，\n提升抗攻击能力，\n但更耗训练。", "relationsNarrative": "Adversarial Example\n它把对抗样本当陪练，让模型学会抗刁钻输入。\n\nData Augmentation\n它像更狠的数据增强，专门加入攻击式变体。\n\nRegularization\n它也有正则化效果，让模型别只记住脆弱捷径。", "relations": { "adversarial-example": { "label": "用…当陪练", "note": "对抗样本是它的主要训练素材。" }, "data-augmentation": { "label": "属于特殊…", "note": "它把恶意扰动也变成训练数据。" }, "regularization": { "label": "提升类似…的鲁棒性", "note": "它通过更难样本限制模型脆弱性。" } } } } }, { "id": "affective-computing", "name": "Affective Computing", "layer": "L4", "era": "1990s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "digital-human" }, { "to": "ai-bias" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "Affective Computing", "factExplain": "Technology that helps machines notice, understand, and respond to human emotions.", "humanExplain": "It is like a very nosy mood coach. You say “I’m fine,” but it notices your tight smile and your “don’t-touch-my-fries” voice.\n\nIt helps AI read feelings and answer with more care. You meet it in companion bots, service chats, learning apps, and cars.", "humanExplainDisplay": "It is like a very ==nosy mood coach==.\nYou say “I’m fine,”\nbut it notices your tight smile\nand your ==“don’t-touch-my-fries” voice==.\n\nIt helps AI read feelings\nand answer with more care.\nYou meet it in companion bots,\nservice chats,\nlearning apps,\nand cars.", "relationsNarrative": "Multimodal AI\nAffective computing often uses voice, faces, and text to guess feelings.\n\nDigital human\nAffective computing can help a Digital human read the room better.\n\nAI-bias\nBad emotion reading can misread some groups of people.\n\nAlignment\nAffective computing affects how AI comforts, persuades, and replies.", "relations": { "multimodal": { "label": "uses … to sense feelings", "note": "It often uses voice, face, and text clues." }, "digital-human": { "label": "helps … read the room", "note": "Emotion reading can make a Digital human feel more responsive." }, "ai-bias": { "label": "can worsen …", "note": "It may misread emotions for some groups of people." }, "alignment": { "label": "shapes … limits", "note": "It must comfort people without pushing or offending them." } } }, "zh": { "fullName": "情感计算", "factExplain": "让机器识别、理解并回应人类情绪的技术方向。", "humanExplain": "它有点像老中医搭脉望闻：你嘴上说“没事”，它还想从语气和表情里看出你到底憋没憋火。\n\n常用于陪伴助手、客服、教育和车载场景，让互动更会察言观色。", "humanExplainDisplay": "它有点像老中医==搭脉望闻==：\n你嘴上说“没事”，\n它还想从语气和表情里\n看出你到底==憋没憋火==。\n\n常用于陪伴助手、\n客服、教育和车载场景，\n让互动更会察言观色。", "relationsNarrative": "Multimodal\n它通常结合语音、表情和文本来判断情绪。\n\nDigital human\n情感计算能让数字人显得更会察言观色。\n\nAI-bias\n情绪识别若有偏差，容易误读不同人群反应。\n\nAlignment\n它会影响 AI 安慰、劝导和回应的分寸边界。", "relations": { "multimodal": { "label": "依赖…感知情绪", "note": "常结合语音、表情和文本线索。" }, "digital-human": { "label": "让…更会察言观色", "note": "情感识别会提升数字人的互动感。" }, "ai-bias": { "label": "容易放大…问题", "note": "不同人群的情绪判断可能失准。" }, "alignment": { "label": "影响…边界", "note": "回应情绪也要避免操控和冒犯。" } } } } }, { "id": "agent-harness", "name": "Agent harness", "layer": "L5", "sublayer": "architecture", "era": "2024", "publishedAt": "2026-05-30T03:10:23.229Z", "relations": [ { "to": "agent" }, { "to": "framework" }, { "to": "human-in-the-loop" }, { "to": "ai-sandbox" } ], "track": "today", "i18n": { "en": { "fullName": "Agent Harness", "factExplain": "A setup for running, testing, and judging an Agent safely.", "humanExplain": "An Agent harness is a test track for a tiny race car. If it swerves toward the snack stand, the guardrail says nope.\n\nBuilders use it while they build an Agent. It tests runs and lowers risk before real work.", "humanExplainDisplay": "An Agent harness is a ==test track==\nfor a tiny race car.\nIf it swerves toward the snack stand,\nthe ==guardrail says nope==.\n\nBuilders use it while they build an Agent.\nIt tests runs and lowers risk\nbefore real work.", "relationsNarrative": "Agent\nAn Agent harness builds the run, test, and review flow around an Agent.\n\nFramework\nAn Agent harness is a framework made for Agent development.\n\nHuman-in-the-loop\nAn Agent harness leaves room for people to check, approve, or take over.\n\nAI sandbox\nAn Agent harness often pairs with an AI sandbox to test behavior safely.", "relations": { "agent": { "label": "controls … runs", "note": "It mainly manages and tests how an Agent runs." }, "framework": { "label": "is a kind of …", "note": "It is a framework built for Agent work." }, "human-in-the-loop": { "label": "makes room for …", "note": "It lets people check and take over at key moments." }, "ai-sandbox": { "label": "often works inside …", "note": "It often uses a safe space to reduce test risk." } } }, "zh": { "fullName": "Agent 运行测试框架", "factExplain": "用于管理、测试和评估 Agent 的运行框架。", "humanExplain": "别看 Agent 会自己忙活，真放出去前还得套个==训练场护栏==：跑偏了能拉回，手欠了能及时==按住别乱点==。\n\n它用于开发和调试 Agent，帮助测试流程并控制风险。", "humanExplainDisplay": "别看 Agent 会自己忙活，\n真放出去前还得套个==训练场护栏==：\n跑偏了能拉回，\n手欠了能及时==按住别乱点==。\n\n它用于开发和调试 Agent，\n帮助测试流程并控制风险。", "relationsNarrative": "Agent\nagent harness 主要围绕 Agent 搭建运行、测试和评估流程。\n\nFramework\nagent harness 本质上是一类专门服务 Agent 开发的框架。\n\nHuman-in-the-loop\nagent harness 往往为人工查看、确认和接管预留位置。\n\nAI sandbox\nagent harness 常与 AI sandbox 搭配，先在隔离环境里验证行为。", "relations": { "agent": { "label": "约束…运行", "note": "它主要用来管理和测试 Agent。" }, "framework": { "label": "属于…一类", "note": "本质上是面向 Agent 的框架。" }, "human-in-the-loop": { "label": "方便…介入", "note": "便于人在关键环节查看和接管。" }, "ai-sandbox": { "label": "常在…中用", "note": "常配合隔离环境降低试错风险。" } } } } }, { "id": "agent-identity", "name": "Agent Identity", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "mcp" }, { "to": "agent-security" }, { "to": "stateful-agent" } ], "track": "today", "i18n": { "en": { "fullName": "Agent Identity", "factExplain": "A system that identifies, authorizes, and tracks an AI agent over time.", "humanExplain": "Agent Identity is an AI worker’s badge. It shows who it is and what it may touch. If trouble starts, the badge leaves footprints.\n\nYou meet it when agents work together or call company tools. It helps control access and trace mistakes.", "humanExplainDisplay": "Agent Identity is an AI worker’s ==badge==.\nIt shows who it is\nand what it may touch.\nIf trouble starts,\nthe badge leaves ==footprints==.\n\nYou meet it when agents work together\nor call company tools.\nIt helps control access\nand trace mistakes.", "relationsNarrative": "Agent\nAgent Identity defines who the Agent is and what it may do.\n\nMCP\nAgent Identity often travels through MCP to outside tools and services.\n\nAgent Security\nAgent Identity is the base for access control and audit trails.\n\nStateful agent\nAgent Identity helps a long-running task keep the same role.", "relations": { "agent": { "label": "gives … an identity", "note": "Without an identity, an Agent is hard to control." }, "mcp": { "label": "passes permissions through …", "note": "Identity often travels through MCP to outside tools." }, "agent-security": { "label": "supports … governance", "note": "Identity is the base for access control and audits." }, "stateful-agent": { "label": "helps … stay continuous", "note": "A steady identity helps long tasks keep the same role." } } }, "zh": { "fullName": "智能体身份", "factExplain": "让智能体可被识别、授权并持续追踪的身份机制。", "humanExplain": "像公司门禁卡加工号：它先把“你是谁、能进哪扇门、出了事找谁”一次说死。\n\n常用于多智能体协作、企业自动化和工具调用，便于控权审计与追责。", "humanExplainDisplay": "像公司门禁卡加工号：\n它先把\n==你是谁==、能进哪扇门，\n==出了事找谁==\n一次说死。\n\n常用于多智能体协作、\n企业自动化和工具调用，\n便于控权审计与追责。", "relationsNarrative": "Agent\n它定义智能体是谁、拥有什么权限。\n\nMCP\n身份常通过协议传给外部工具和服务。\n\nAgent Security\n身份是权限控制、审计追踪的基础。\n\nStateful agent\n持续身份能让长期任务保持同一角色。", "relations": { "agent": { "label": "定义…的身份", "note": "没有身份，智能体就难被管住。" }, "mcp": { "label": "配合…传递权限", "note": "身份常随协议传到外部工具。" }, "agent-security": { "label": "支撑…治理", "note": "身份是权限控制和审计基础。" }, "stateful-agent": { "label": "帮助…保持连续性", "note": "持续身份让长期任务不断档。" } } } } }, { "id": "agent-internet", "name": "Agent Internet", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "agent-identity" }, { "to": "agent-tool-discovery" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "Agent Internet", "factExplain": "A network system where Agents find, talk, and work with each other online.", "humanExplain": "Agent Internet is like a group chat for AI helpers. The real ones pass jobs around and boot the fake intern.\n\nIt lets Agents find each other across apps. IDs and safety rules keep it from becoming robot spam.", "humanExplainDisplay": "Agent Internet is like a ==group chat==\nfor AI helpers.\nThe real ones ==pass jobs around==\nand boot the fake intern.\n\nIt lets Agents find each other\nacross apps.\nIDs and safety rules keep it\nfrom becoming robot spam.", "relationsNarrative": "Agent\nAgent Internet connects many Agents into a bigger working system.\n\nAgent Identity\nAgent Identity shows who an Agent is and what it may do.\n\nAgent Tool Discovery\nAgent Tool Discovery helps an Agent find useful services and partners.\n\nAgent Security\nAgent Security helps stop fake Agents, abuse, and attacks.", "relations": { "agent": { "label": "connects many …", "note": "Many Agents can work together through the network." }, "agent-identity": { "label": "checks trust with …", "note": "Identity lets an Agent be trusted and held responsible." }, "agent-tool-discovery": { "label": "finds services through …", "note": "Discovery helps an Agent find the right service or partner." }, "agent-security": { "label": "must guard against … risks", "note": "Open networks give attackers more places to try." } } }, "zh": { "fullName": "智能体互联网", "factExplain": "让智能体在线发现、通信并协作的网络体系。", "humanExplain": "智能体互联网像给各路 AI 开江湖镖局：接单、递信、找帮手，也得防冒牌侠客。\n\n用于跨应用协作和服务调用，身份、协议、安全是底座。", "humanExplainDisplay": "智能体互联网\n像给各路 AI 开==江湖镖局==：\n接单、递信、找帮手，\n也得防==冒牌侠客==。\n\n用于跨应用协作和服务调用，\n身份、协议、安全是底座。", "relationsNarrative": "Agent\n单个 Agent 通过网络互联，协作成更大体系。\n\nAgent Identity\n身份让智能体可识别、可授权、可追责。\n\nAgent Tool Discovery\n发现机制让智能体找到可用服务和伙伴。\n\nAgent Security\n开放互联会放大冒充、滥用和攻击风险。", "relations": { "agent": { "label": "连接多个…", "note": "单个 Agent 通过网络协作成体系。" }, "agent-identity": { "label": "依赖…识别身份", "note": "身份让代理可被信任和追责。" }, "agent-tool-discovery": { "label": "需要…发现服务", "note": "发现机制决定代理能找到谁。" }, "agent-security": { "label": "必须防住…风险", "note": "开放互联会放大攻击面。" } } } } }, { "id": "agent-memory", "name": "Memory", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "stateful-agent" }, { "to": "rag" }, { "to": "vector-db" } ], "track": "today", "i18n": { "en": { "fullName": "Agent memory", "factExplain": "A way for an Agent to save and reuse information across chats.", "humanExplain": "It is the sticky note your AI keeps on the fridge. It remembers your pizza order and your pet peeves, so you do not start from zero.\n\nYou meet it in long jobs and personal helpers. It lets the Agent keep going across many steps.", "humanExplainDisplay": "It is the ==sticky note==\nyour AI keeps on the fridge.\nIt remembers your pizza order\nand your ==pet peeves==,\nso you do not start from zero.\n\nYou meet it in long jobs\nand personal helpers.\nIt lets the Agent keep going\nacross many steps.", "relationsNarrative": "Agent\nMemory lets an Agent keep preferences, tasks, and context across turns.\n\nStateful agent\nA Stateful agent often uses memory to keep its actions connected.\n\nRAG\nLong-term memory often uses RAG to find old information when needed.\n\nVector-db\nMany memories are stored as vectors in a Vector-db for easy search.", "relations": { "agent": { "label": "helps … remember", "note": "Without memory, many Agents forget after each chat." }, "stateful-agent": { "label": "forms the base for …", "note": "Memory helps a Stateful agent keep going from turn to turn." }, "rag": { "label": "can use … to recall", "note": "RAG can pull old memory back when needed." }, "vector-db": { "label": "is often stored in …", "note": "Many memories become vectors so a Vector-db can find them fast." } } }, "zh": { "fullName": "Agent memory｜智能体记忆", "factExplain": "让智能体跨轮次保存并调用信息的机制。", "humanExplain": "终于不用每次从头交代，它像楼下常去的煎饼摊大姐：你不要香菜、多加蛋全记着，下次直接给你安排上。\n\n适合长期任务、个性化助手和多步流程，让智能体越用越顺手。", "humanExplainDisplay": "终于不用每次从头交代，\n它像==楼下常去的煎饼摊大姐==：\n你不要香菜、多加蛋全记着，\n下次直接==给你安排上==。\n\n适合长期任务、个性化助手和多步流程，\n让智能体越用越顺手。", "relationsNarrative": "Agent\n记忆让 Agent 能跨轮次保留偏好、任务和上下文。\n\nStateful agent\n状态型智能体通常依赖记忆维持连续行为。\n\nRAG\n长期记忆常结合 RAG，在需要时再检索回来。\n\nVector-db\n很多记忆会以向量形式存进 Vector-db 便于查找。", "relations": { "agent": { "label": "让…记住事", "note": "没有记忆，很多智能体只能现聊现忘。" }, "stateful-agent": { "label": "构成…基础", "note": "状态型智能体通常靠记忆维持连续性。" }, "rag": { "label": "可借…取回", "note": "长期记忆常要靠检索把信息找回来。" }, "vector-db": { "label": "常存进…", "note": "很多记忆会向量化后存入数据库。" } } } } }, { "id": "agent-native-tools", "name": "Agent-native Tools", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "mcp" }, { "to": "agent-tool-discovery" } ], "track": "today", "i18n": { "en": { "fullName": "Agent-native Tools", "factExplain": "Tool interfaces built for an AI Agent to call directly.", "humanExplain": "Normal apps make an AI tap around like a cat on an iPad. Agent-native Tools give it big labeled buttons.\n\nThey let an Agent skip pages and call the tool directly. You see them in office bots and coding bots.", "humanExplainDisplay": "Normal apps make an AI tap around\nlike a ==cat on an iPad==.\nAgent-native Tools give it\n==big labeled buttons==.\n\nThey let an Agent skip pages\nand call the tool directly.\nYou see them in office bots\nand coding bots.", "relationsNarrative": "Agent\nAgent-native Tools give the Agent a direct entrance to get work done.\n\nFunction-calling\nFunction-calling turns the Agent’s intent into a real tool action.\n\nMCP\nMCP often connects these tools into an Agent system.\n\nAgent Tool Discovery\nAgent Tool Discovery helps the Agent pick the right tool.", "relations": { "agent": { "label": "lets … use tools directly", "note": "The tool becomes an interface the Agent can read and use." }, "function-call": { "label": "runs actions through …", "note": "Function-calling turns the Agent’s intent into a real tool action." }, "mcp": { "label": "can plug into …", "note": "MCP connects many tools to an Agent in one common way." }, "agent-tool-discovery": { "label": "works with … to find tools", "note": "Discovery helps the Agent know which tool to use." } } }, "zh": { "fullName": "Agent-native Tools（代理原生工具）", "factExplain": "为 AI 代理设计的可调用工具接口。", "humanExplain": "普通 App 让 AI 像隔着手套戳屏幕；代理原生工具直接给它一排贴好标签的大按钮。\n\n让代理少绕页面，直接调接口干活，办自动化办公和开发。", "humanExplainDisplay": "普通 App 让 AI 像\n==隔着手套戳屏幕==；\n代理原生工具直接给它\n==一排贴标签的大按钮==。\n\n让代理少绕页面，\n直接调接口干活，\n办自动化办公和开发。", "relationsNarrative": "Agent\nAgent-native Tools 是 Agent 可直接调用的办事入口。\n\nFunction-calling\nFunction-calling 把代理意图转成具体工具操作。\n\nMCP\nMCP 常用来把这些工具统一接入代理系统。\n\nAgent Tool Discovery\nAgent Tool Discovery 让代理知道该选哪个工具。", "relations": { "agent": { "label": "让…直接使用", "note": "工具变成 Agent 能读会用的接口。" }, "function-call": { "label": "通过…执行动作", "note": "函数调用把意图落到具体操作。" }, "mcp": { "label": "可接入…", "note": "MCP 负责把工具统一接给代理。" }, "agent-tool-discovery": { "label": "配合…找工具", "note": "发现机制让代理知道该用哪一个。" } } } } }, { "id": "agent-observability", "name": "Agent observability", "layer": "L5", "sublayer": "architecture", "era": "2023", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "llmops" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "Agent Observability", "factExplain": "A monitoring method for recording and checking an Agent’s steps.", "humanExplain": "Agent observability is a dashcam for your AI helper. If it crashes the shopping cart, you can replay the wobble.\n\nYou meet it when Agents run real work. It tracks plans and tool calls, so teams can find failures and audit actions.", "humanExplainDisplay": "Agent observability is a ==dashcam==\nfor your AI helper.\nIf it crashes the shopping cart,\nyou can ==replay the wobble==.\n\nYou meet it when Agents run real work.\nIt tracks plans and tool calls,\nso teams can find failures and audit actions.", "relationsNarrative": "Agent\nAgent observability records the Agent’s plans, tool calls, and results.\n\nFunction-calling\nFunction-calling logs show which step failed.\n\nLLMOps\nAgent observability adds the Agent path to post-launch monitoring.\n\nAgent Security\nAgent observability gives audit clues and helps spot strange actions.", "relations": { "agent": { "label": "records … behavior", "note": "Observability leaves a trail for every Agent step." }, "function-call": { "label": "tracks … results", "note": "Tool call logs show which step got stuck." }, "llmops": { "label": "adds to … monitoring", "note": "LLMOps must watch the Agent path after launch." }, "agent-security": { "label": "supports … audits", "note": "Action trails help spot off-limits access and odd moves." } } }, "zh": { "fullName": "Agent 可观测性", "factExplain": "记录并分析 Agent 行为轨迹的监控方法。", "humanExplain": "Agent 可观测性就是棋局复盘：不只看输赢，还看哪步昏招送车。\n\n它追踪计划和工具调用，定位失败并支撑审计。", "humanExplainDisplay": "Agent 可观测性就是\n==棋局复盘==：\n不只看输赢，\n还看哪步==昏招送车==。\n\n它追踪计划和工具调用，\n定位失败，\n并支撑审计。", "relationsNarrative": "Agent\n它记录 Agent 的计划、工具调用和结果。\n\nFunction-calling\n工具调用日志帮助定位失败发生在哪一步。\n\nLLMOps\n它把智能体链路纳入上线后的监控。\n\nAgent Security\n可观测性提供审计线索，发现异常行动。", "relations": { "agent": { "label": "记录…的行为", "note": "可观测性把 Agent 的每步行动留痕。" }, "function-call": { "label": "追踪…结果", "note": "工具调用日志能定位卡在哪一步。" }, "llmops": { "label": "纳入…监控", "note": "上线后监控需覆盖智能体链路。" }, "agent-security": { "label": "支撑…审计", "note": "行为留痕能发现越权和异常操作。" } } } } }, { "id": "agent-security", "name": "Agent Security", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "prompt-injection" }, { "to": "computer-use" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI Agent Security", "factExplain": "Rules to stop an AI agent from being tricked, misused, or going too far.", "humanExplain": "Think of a helpful intern with a badge and keycard. Without rules, a fake boss might get them to sign for 80 pizzas.\n\nYou need it when agents use the web or run company tasks. It stops bad access and wrong moves.", "humanExplainDisplay": "Think of a helpful intern\nwith a ==badge== and ==keycard==.\nWithout rules,\na fake boss might get them\nto sign for 80 pizzas.\n\nYou need it when agents use the web\nor run company tasks.\nIt stops bad access\nand wrong moves.", "relationsNarrative": "Agent\nAgent security matters more as an Agent gets more freedom to act.\n\nPrompt injection\nPrompt injection can steer the agent off task and into unsafe actions.\n\nComputer use\nComputer use makes wrong clicks and wrong sends matter more.\n\nHuman-in-the-loop\nHuman-in-the-loop puts a person in charge of risky final choices.", "relations": { "agent": { "label": "protects … actions", "note": "The more an Agent can do, the more it needs safe limits." }, "prompt-injection": { "label": "blocks … attacks", "note": "Prompt injection can trick the agent away from its real task." }, "computer-use": { "label": "limits … actions", "note": "Mouse and keyboard control makes mistakes more serious." }, "human-in-the-loop": { "label": "adds … backup", "note": "A human often approves high-risk steps before they happen." } } }, "zh": { "fullName": "AI 代理安全", "factExplain": "保护代理执行任务时不被诱导、滥用或越权。", "humanExplain": "像给能自己跑腿的打工人上工牌、开权限、盯流程；不然别人一忽悠，它真敢替你乱签字。\n\n常用于电脑操作、联网调用和企业自动化，重点防越权与误执行。", "humanExplainDisplay": "像给能自己跑腿的打工人\n上==工牌、开权限、盯流程==；\n不然别人一忽悠，\n它真敢替你==乱签字==。\n\n常用于电脑操作、联网调用\n和企业自动化，\n重点防越权与误执行。", "relationsNarrative": "Agent\n代理越能自主做事，安全问题就越关键。\n\nPrompt Injection\n提示词注入会把它带偏，诱导泄密或越权。\n\nComputer Use\n一旦能操作电脑，误点误发的后果更直接。\n\nHuman-in-the-loop\n高风险任务常靠人工审批做最后保险。", "relations": { "agent": { "label": "保护…执行", "note": "代理越能行动，越需要安全边界。" }, "prompt-injection": { "label": "防范…攻击", "note": "提示词注入会诱导它偏离原任务。" }, "computer-use": { "label": "约束…操作", "note": "能点鼠标敲键盘时风险会放大。" }, "human-in-the-loop": { "label": "引入…兜底", "note": "高风险步骤常要人来最终拍板。" } } } } }, { "id": "agent-tool-discovery", "name": "Agent Tool Discovery", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "mcp" }, { "to": "prompt-injection" } ], "track": "today", "i18n": { "en": { "fullName": "Agent Tool Discovery", "factExplain": "A way for an Agent to find and choose tools it can use.", "humanExplain": "Agent Tool Discovery is like opening a mystery toolbox. First, find the mop, not the leaf blower.\n\nYou see it in multi-tool Agents and company workflows. It chooses the right API, so the task has a chance.", "humanExplainDisplay": "Agent Tool Discovery is like opening a ==mystery toolbox==.\nFirst, find the ==mop==,\nnot the leaf blower.\n\nYou see it in multi-tool Agents\nand company workflows.\nIt chooses the right API,\nso the task has a chance.", "relationsNarrative": "Agent\nTool discovery helps an Agent find its tools before it acts.\n\nFunction-calling\nTool discovery usually happens before Function-calling starts.\n\nMCP\nMCP often shows the Agent a list of tools and notes.\n\nPrompt injection\nPrompt injection can trick it into choosing the wrong tool.", "relations": { "agent": { "label": "helps … find tools", "note": "It lets an Agent see available tools before it acts." }, "function-call": { "label": "comes before …", "note": "The Agent should pick the right tool before it calls it." }, "mcp": { "label": "often connects through …", "note": "MCP can show the Agent a menu of tools." }, "prompt-injection": { "label": "can be tricked by …", "note": "Bad instructions can push it toward the wrong tool." } } }, "zh": { "fullName": "Agent 工具发现", "factExplain": "让代理识别并选择可用工具的机制。", "humanExplain": "像武侠里先摸清各门派路数：谁会轻功，谁擅医术，别一着急就把求救信递错人。\n\n常见于多工具代理和企业流程，决定该调哪个接口，也影响任务成功率。", "humanExplainDisplay": "像武侠里先摸清\n各门派==路数==：\n谁会轻功，谁擅医术，\n别把求救信==递错人==。\n\n常见于多工具代理\n和企业流程，\n决定该调哪个接口，\n也影响任务成功率。", "relationsNarrative": "Agent\n它帮代理先识别可用工具，再决定怎么行动。\n\nFunction-call\n工具发现通常发生在发起工具调用之前。\n\nMCP\nMCP 常用来向代理暴露工具清单和说明。\n\nPrompt injection\n恶意提示可能误导它选择不该用的工具。", "relations": { "agent": { "label": "帮…找工具", "note": "让代理先识别可用能力再行动。" }, "function-call": { "label": "为…做前置", "note": "先选对工具，后面才好发起调用。" }, "mcp": { "label": "常通过…接入", "note": "工具目录常借 MCP 暴露给代理。" }, "prompt-injection": { "label": "会受…干扰", "note": "恶意说明可能诱导它选错工具。" } } } } }, { "id": "agent", "name": "Agent", "aliases": [ "智能体" ], "layer": "L5", "sublayer": "architecture", "era": "2023", "publishedAt": "2026-05-23T09:40:00Z", "relations": [ { "to": "llm" }, { "to": "function-call" }, { "to": "mcp" }, { "to": "framework" } ], "i18n": { "en": { "fullName": "AI Agent", "factExplain": "AI software for planning a task and doing the steps on its own.", "humanExplain": "A chatbot is the kid who talks about chores. An Agent grabs the broom.\n\nYou meet it in tools that sort email or book trips. It helps, but check its work like a new intern.", "humanExplainDisplay": "A chatbot is the kid\nwho ==talks about chores==.\nAn Agent ==grabs the broom==.\n\nYou meet it in tools\nthat sort email or book trips.\nIt helps,\nbut check its work\nlike a new intern.", "relationsNarrative": "LLM\nThe Agent uses an LLM to understand the goal and plan the next move.\n\nFunction-calling\nFunction-call lets the Agent turn a plan into tool actions.\n\nMCP\nMCP gives the Agent one standard way to connect tools and data.\n\nFramework\nA Framework organizes the Agent parts into a working app.", "relations": { "llm": { "label": "uses … as its brain", "note": "The LLM helps the Agent understand the goal and plan." }, "function-call": { "label": "calls tools with …", "note": "Function-call turns the Agent plan into tool actions." }, "mcp": { "label": "connects tools through …", "note": "MCP gives the Agent one way to reach tools and data." }, "framework": { "label": "is built with …", "note": "A Framework helps package the Agent into a working app." } } }, "zh": { "fullName": "AI 代理", "factExplain": "一种能够自主规划并执行任务的 AI 系统。", "humanExplain": "AI 代理像请了个电子跑腿小哥，不只回你“收到”，还会真的去办事。\n\n它能整理邮件、订会议、查资料，但复杂任务仍要人盯着方向。", "humanExplainDisplay": "AI 代理像请了个==电子跑腿小哥==，\n不只回你“收到”，\n还会==真的去办事==。\n\n它能整理邮件、订会议、查资料，\n但复杂任务仍要人盯着方向。", "relationsNarrative": "LLM\nAgent 依赖 LLM 理解目标，并规划下一步行动。\n\nFunction-calling\nFunction-call 让 Agent 能把计划转化为工具调用。\n\nMCP\nMCP 为 Agent 连接工具和数据源提供统一协议。\n\nFramework\nFramework 将 Agent 的模型、工具和流程组织成应用。", "relations": { "llm": { "label": "用…作为大脑" }, "function-call": { "label": "用…调用工具" }, "mcp": { "label": "用…接工具" } } } } }, { "id": "agentic-coding", "name": "Agentic coding", "layer": "L4", "era": "2025", "publishedAt": "2026-05-31T00:57:33.086Z", "relations": [ { "to": "agent" }, { "to": "cursor" }, { "to": "vibe-coding" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "Agentic coding", "factExplain": "A way for AI to handle coding tasks on its own.", "humanExplain": "Agentic coding is like a junior coder with too much coffee in your laptop. You give one request, and it starts poking through the project.\n\nYou meet it in code helpers and IDEs like Cursor. It can fix bugs or sketch prototypes, but a human still approves big changes.", "humanExplainDisplay": "Agentic coding is like a ==junior coder==\nwith too much coffee in your laptop.\nYou give one request,\nand it starts ==poking through the project==.\n\nYou meet it in code helpers\nand IDEs like Cursor.\nIt can fix bugs or sketch prototypes,\nbut a human still approves big changes.", "relationsNarrative": "Agent\nAgentic coding is an Agent used for software development.\n\nCursor\nCursor and similar tools make agentic coding part of everyday work.\n\nVibe-coding\nVibe-coding is more like verbal steering. Agentic coding is more hands-on.\n\nHuman-in-the-loop\nImportant code changes still need a human to review and approve them.", "relations": { "agent": { "label": "uses … for coding", "note": "It is an Agent working inside software development." }, "cursor": { "label": "shows up in …", "note": "Tools like Cursor make it a daily coding flow." }, "vibe-coding": { "label": "does more than …", "note": "Vibe-coding gives direction. Agentic coding does the work." }, "human-in-the-loop": { "label": "needs … approval", "note": "People still review and approve important code changes." } } }, "zh": { "fullName": "代理式编程", "factExplain": "让 AI 像代理一样自主完成编程任务的方式。", "humanExplain": "你发一句需求，它就自己拆任务、改文件、跑测试，活像工位上来了个不用催的实习生。\n\n常见于代码助手和 IDE，适合修 Bug、搭原型；关键改动仍要人拍板。", "humanExplainDisplay": "你发一句需求，\n它就自己拆任务、改文件、跑测试，\n活像工位上来了个\n==不用催的实习生==。\n\n常见于代码助手和 IDE，\n适合修 Bug、搭原型；\n关键改动仍要人拍板。", "relationsNarrative": "Agent\nAgentic coding 是 Agent 在软件开发场景中的具体用法。\n\nCursor\nCursor 等工具把 Agentic coding 做成日常开发体验。\n\nVibe-coding\nVibe-coding 更像口头指挥，Agentic coding 更强调自主执行。\n\nHuman-in-the-loop\n涉及关键代码改动时，通常仍需要人类审核把关。", "relations": { "agent": { "label": "是…在编程中的用法", "note": "本质是 Agent 落到软件开发场景。" }, "cursor": { "label": "常见于…这类工具", "note": "不少 IDE 产品把它做成核心体验。" }, "vibe-coding": { "label": "比…更强调执行", "note": "不只提建议，还会动手改代码。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "关键改动仍要人审核与拍板。" } } } } }, { "id": "agentic-commerce", "name": "Agentic commerce", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "computer-use" }, { "to": "function-call" }, { "to": "permission-fatigue" } ], "track": "today", "i18n": { "en": { "fullName": "Agentic commerce", "factExplain": "A way for AI agents to compare, choose, and finish purchases for people.", "humanExplain": "It is a bargain-hunter friend loose in Target. You are still picking a cart. It already found the deal and paid.\n\nIt can compare prices and buy things. It can book trips. It can help companies purchase supplies. It needs clear permission, rules, and safety checks first.", "humanExplainDisplay": "It is a ==bargain-hunter friend== loose in Target.\nYou are still picking a cart.\nIt already ==found the deal== and paid.\n\nIt can compare prices and buy things.\nIt can book trips.\nIt can help companies purchase supplies.\nIt needs clear permission,\nrules,\nand safety checks first.", "relationsNarrative": "Agent\nAgentic commerce uses an Agent to make shopping and buying choices for people.\n\nComputer use\nComputer use lets it operate websites, apps, and forms directly.\n\nFunction-calling\nFunction-call connects it to stores, payments, and order systems.\n\nPermission fatigue\nToo many permission steps can make users resist handing over control.", "relations": { "agent": { "label": "brings … into buying", "note": "Agentic commerce puts Agents inside shopping and purchasing." }, "computer-use": { "label": "buys through …", "note": "Computer use lets it click websites and apps for you." }, "function-call": { "label": "connects stores with …", "note": "Function-call lets it check prices, place orders, and track them." }, "permission-fatigue": { "label": "can cause …", "note": "Too many approvals can make users tired of giving permission." } } }, "zh": { "fullName": "代理式商业", "factExplain": "由 AI 代理发起、比较并完成交易的商业形态。", "humanExplain": "你还在购物车里来回横跳，它已经像夜市里会砍价的老江湖：货比三摊、嘴皮一翻，顺手就替你拿下。\n\n用于比价购买、订票订酒店和企业采购，前提是权限、规则和风控配好。", "humanExplainDisplay": "你还在购物车里来回横跳，\n它已经像夜市里会砍价的==老江湖==：\n货比三摊、嘴皮一翻，\n顺手就替你==拿下==。\n\n用于比价购买、\n订票订酒店和企业采购，\n前提是权限、规则和风控配好。", "relationsNarrative": "Agent\n它本质上是 Agent 代替人做购物与采购决策。\n\nComputer use\nComputer use 让它能直接操作网页、App 和表单。\n\nFunction-calling\nFunction-call 让它连接商家、支付和订单系统。\n\nPermission fatigue\n授权步骤太多时，用户容易对放权产生抗拒。", "relations": { "agent": { "label": "把…搬进交易", "note": "本质是 Agent 进入购物与采购流程。" }, "computer-use": { "label": "靠…代点代买", "note": "能操作网页和应用，才可能代你下单。" }, "function-call": { "label": "用…连商家系统", "note": "通过接口查价、下单、支付与查询。" }, "permission-fatigue": { "label": "容易带来…", "note": "授权事项一多，用户会越来越麻。" } } } } }, { "id": "agentic-search", "name": "Agentic search", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "rag" }, { "to": "information-retrieval" }, { "to": "prompt-injection" } ], "track": "today", "i18n": { "en": { "fullName": "Agentic search", "factExplain": "AI plans steps, searches many times, then combines what it finds.", "humanExplain": "Agentic search is not a stack of blue links. It is a very online friend with a clipboard, checking the weird one-star reviews for you.\n\nIt helps with hard questions and research. It saves time, but bad pages can still fool it.", "humanExplainDisplay": "Agentic search is not\na ==stack of blue links==.\nIt is a ==very online friend with a clipboard==,\nchecking the weird one-star reviews for you.\n\nIt helps with hard questions and research.\nIt saves time,\nbut bad pages can still fool it.", "relationsNarrative": "Agent\nAgentic search is an Agent used for search tasks.\n\nRAG\nAgentic search is more active than RAG and plans its own search steps.\n\nIR\nAgentic search builds on IR because search systems find the information.\n\nPrompt injection\nAgentic search reads web pages, so prompt injection can lead it off track.", "relations": { "agent": { "label": "is search form of …", "note": "It turns an Agent into a search doer, not just an answer box." }, "rag": { "label": "is more active than …", "note": "It plans several searches instead of only fetching source text." }, "information-retrieval": { "label": "builds on …", "note": "IR gives it the way to find useful pages." }, "prompt-injection": { "label": "can be misled by …", "note": "Bad pages can plant instructions and push it off track." } } }, "zh": { "fullName": "代理式搜索", "factExplain": "AI 自主规划、多步检索并整合答案的搜索方式。", "humanExplain": "不是甩你一堆链接自己翻，更像找了个会比价会看差评的代购：跑遍全网，最后把结论拎给你。\n\n适合复杂检索和调研；省时间，但也可能被坏信息带偏。", "humanExplainDisplay": "不是甩你一堆链接自己翻，\n更像找了个\n==会比价会看差评==的代购：\n跑遍全网，最后把\n==结论拎给你==。\n\n适合复杂检索和调研；\n省时间，但也可能被坏信息带偏。", "relationsNarrative": "Agent\n它是 Agent 在搜索场景里的典型落地形态。\n\nRAG\n它比 RAG 更主动，会自己拆解问题并多步查找。\n\nInformation Retrieval\n它建立在检索系统之上，由检索提供信息入口。\n\nPrompt Injection\n它会读网页内容，因此更容易受恶意提示干扰。", "relations": { "agent": { "label": "是…的搜索形态", "note": "把搜索从问答升级成主动执行。" }, "rag": { "label": "比…更主动", "note": "不只取资料，还会规划多步搜索。" }, "information-retrieval": { "label": "建立在…之上", "note": "底层仍靠检索系统找信息。" }, "prompt-injection": { "label": "容易受…干扰", "note": "恶意网页可能诱导它跑偏。" } } } } }, { "id": "agi", "name": "AGI", "layer": "L6", "era": "2000s", "publishedAt": "2026-05-23T11:35:00Z", "relations": [ { "to": "superintelligence" }, { "to": "alignment" }, { "to": "emergence" }, { "to": "singularity" } ], "i18n": { "en": { "fullName": "Artificial General Intelligence", "factExplain": "A still-theoretical AI for learning and doing many kinds of work like a person.", "humanExplain": "AGI is the rumored new kid: straight A’s in every class. At lunch, everyone argues if the kid even goes here.\n\nIt turns AI from one-use tools toward a general mind. That brings safety fights, rule fights, and control fears.", "humanExplainDisplay": "AGI is the ==rumored new kid==:\n==straight A’s in every class==.\nAt lunch,\neveryone argues\nif the kid even goes here.\n\nIt turns AI from one-use tools\ntoward a general mind.\nThat brings safety fights,\nrule fights,\nand control fears.", "relationsNarrative": "Superintelligence\nSuperintelligence is the extreme idea of AGI getting far stronger.\n\nAlignment\nThe stronger AGI gets, the more Alignment matters for control.\n\nEmergence\nEmergence says AGI-like skills may appear in jumps, not in a straight line.\n\nSingularity\nSingularity describes AGI rushing past the edge of human control.", "relations": { "superintelligence": { "label": "could lead to …", "note": "Superintelligence is the extreme idea of AGI getting far stronger." }, "alignment": { "label": "needs …", "note": "Stronger AGI makes control and goals much more important." }, "emergence": { "label": "may appear through …", "note": "Emergence may explain AGI-like skills showing up in jumps." }, "singularity": { "label": "could trigger …", "note": "Singularity is the fear of AGI racing past human control." } } }, "zh": { "fullName": "通用人工智能", "factExplain": "理论上能像人一样跨领域学习和完成任务的 AI。", "humanExplain": "它像小区门口的万能师傅，修水管、写作业、做PPT，啥活都敢接，还不挑单。\n\n它代表 AI 的长期目标，也牵出对齐、安全和监管问题。", "humanExplainDisplay": "它像==小区门口的万能师傅==，\n修水管、写作业、做PPT，\n==啥活都敢接==，还不挑单。\n\n它代表 AI 的长期目标，\n也牵出对齐、安全和监管问题。", "relationsNarrative": "Superintelligence\nSuperintelligence 是 AGI 继续增强后的极端能力假设。\n\nAlignment\nAGI 能力越强，Alignment 越关系到系统是否可控。\n\nEmergence\nEmergence 解释 AGI 能力可能以非线性方式出现。\n\nSingularity\nSingularity 描述 AGI 快速越过人类控制边界的情景。", "relations": { "superintelligence": { "label": "可能通向…" }, "alignment": { "label": "需要…" }, "emergence": { "label": "与…相关" } } } } }, { "id": "ai-abstention", "name": "AI Abstention", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "hallucination" }, { "to": "model-uncertainty" }, { "to": "human-in-the-loop" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "AI Abstention", "factExplain": "A safety rule for AI to refuse unsure or risky questions.", "humanExplain": "AI Abstention is a school nurse staring at mystery pills. No label, no guessing, no random jelly bean for you.\n\nYou meet it in high-stakes places, like health or law. It cuts bad advice. The trade-off is blunt: some questions get no answer.", "humanExplainDisplay": "AI Abstention is a ==school nurse==\nstaring at ==mystery pills==.\nNo label.\nNo guessing.\nNo ==random jelly bean== for you.\n\nYou meet it in high-stakes places,\nlike health or law.\nIt cuts bad advice.\nThe trade-off is blunt:\nsome questions get no answer.", "relationsNarrative": "Hallucination\nAI Abstention reduces Hallucination by refusing to make things up.\n\nModel uncertainty\nAI Abstention uses Model uncertainty to decide if it should answer.\n\nHuman-in-the-loop\nAI Abstention can send refused questions to a human.\n\nAlignment\nAI Abstention supports Alignment by keeping the AI inside safe limits.", "relations": { "hallucination": { "label": "cuts down on …", "note": "Staying quiet when unsure means fewer made-up answers." }, "model-uncertainty": { "label": "uses … to decide", "note": "The AI checks its confidence before it answers." }, "human-in-the-loop": { "label": "hands hard cases to …", "note": "If the AI refuses, a person may take over." }, "alignment": { "label": "supports …", "note": "Abstention helps the AI stay careful instead of bluffing." } } }, "zh": { "fullName": "AI 弃答", "factExplain": "AI 在不确定或高风险时选择不回答的机制。", "humanExplain": "真没把握时，它像老中医不给你乱开方：脉没摸准就先不开药，总比瞎抓一把让人更遭罪。\n\n常用于高风险场景，减少误导；代价是有些问题会直接不答。", "humanExplainDisplay": "真没把握时，它像==老中医==\n不给你乱开方：脉没摸准\n就先不开药，总比\n==瞎抓一把==让人更遭罪。\n\n常用于高风险场景，减少误导；\n代价是有些问题会直接不答。", "relationsNarrative": "Hallucination\n弃答通过“不乱答”，来减少幻觉带来的误导。\n\nModel uncertainty\n它通常依据不确定性判断是否该回答。\n\nHuman-in-the-loop\n当模型选择弃答时，任务常会转给人工处理。\n\nAlignment\n弃答是让模型更稳妥、更守边界的一种办法。", "relations": { "hallucination": { "label": "用弃答压住…", "note": "不确定时闭嘴，能少编错答案。" }, "model-uncertainty": { "label": "依据…决定", "note": "先判断把握够不够，再决定答不答。" }, "human-in-the-loop": { "label": "把难题交给…", "note": "模型不答时，常转人工兜底处理。" }, "alignment": { "label": "属于…手段", "note": "让模型更稳妥，而非逞强乱答。" } } } } }, { "id": "ai-academic-integrity", "name": "AI Academic Integrity", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "ai-detector" }, { "to": "ai-proctoring" }, { "to": "ai-literacy-ai" }, { "to": "ai-assisted-research" } ], "track": "today", "i18n": { "en": { "fullName": "AI Academic Integrity", "factExplain": "Honesty rules for using AI in school work and writing.", "humanExplain": "AI Academic Integrity is the class rule for calculators with a mouth. Use the helper, but do not let it write your whole paper.\n\nYou meet it in class rules, paper credit, and appeals. The big point is clear: what you did, and what AI did.", "humanExplainDisplay": "AI Academic Integrity is the ==class rule==\nfor ==calculators with a mouth==.\nUse the helper,\nbut do not let it write your whole paper.\n\nYou meet it in class rules,\npaper credit,\nand appeals.\nThe big point is clear:\nwhat you did,\nand what AI did.", "relationsNarrative": "AI detector\nAI detectors can flag odd work, but people still need to judge.\n\nAI Proctoring\nAI Proctoring watches exam behavior, while integrity rules set the limits.\n\nAI Literacy\nAI Literacy helps students know when to say they used AI.\n\nAI-assisted Research\nAI-assisted Research can save time, but human and AI roles must be clear.", "relations": { "ai-detector": { "label": "uses … for clues", "note": "An AI detector gives clues, not final proof." }, "ai-proctoring": { "label": "uses … to watch exams", "note": "Proctoring watches behavior, but rules define honest use." }, "ai-literacy-ai": { "label": "sets rules through …", "note": "AI Literacy helps students know when to say they used AI." }, "ai-assisted-research": { "label": "sets rules for …", "note": "Research can use AI, but the roles must be clear." } } }, "zh": { "fullName": "AI 学术诚信", "factExplain": "一套规范 AI 辅助学习与写作的诚信原则。", "humanExplain": "AI 学术诚信像开卷考：能翻书，别把 AI 代写当自己满分作文。\n\n用于课程规则、论文署名和申诉，关键是说清人机分工。", "humanExplainDisplay": "AI 学术诚信像==开卷考==：\n能翻书，\n别把 AI 代写\n当==自己满分作文==。\n\n用于课程规则、论文署名\n和申诉，\n关键是说清人机分工。", "relationsNarrative": "AI Detector\n检测器能辅助发现异常，但不能替代学术判断。\n\nAI Proctoring\n远程监考关注考试行为，诚信规则管使用边界。\n\nAI Literacy\nAI 素养让学生知道哪些用法该标注、该避开。\n\nAI-assisted Research\nAI 辅助研究可以提效，但人机贡献要说清。", "relations": { "ai-detector": { "label": "用…辅助判断", "note": "检测器只能给线索，不能当铁证。" }, "ai-proctoring": { "label": "用…监督考试", "note": "监考工具管行为，诚信还要靠规则。" }, "ai-literacy-ai": { "label": "靠…提前立规矩", "note": "懂 AI 边界，才知道何时该声明。" }, "ai-assisted-research": { "label": "规范…的使用", "note": "研究可用 AI，但贡献要说清。" } } } } }, { "id": "ai-accelerated-prototyping", "name": "AI Prototyping", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "copilot" }, { "to": "agentic-coding" }, { "to": "vibe-coding" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI-Accelerated Prototyping", "factExplain": "A way to use AI to build product prototypes faster.", "humanExplain": "You get an app idea over cereal. AI slides over a tiny free sample, with buttons.\n\nTeams use it to test ideas fast. They can choose early: keep going, or stop.", "humanExplainDisplay": "You get an app idea over cereal.\nAI slides over a ==tiny free sample==,\nwith ==buttons==.\n\nTeams use it to test ideas fast.\nThey can choose early:\nkeep going, or stop.", "relationsNarrative": "Copilot\nAI prototyping often uses Copilot to make the first version fast.\n\nAgentic coding\nAgentic coding is a more hands-off step beyond AI prototyping.\n\nVibe-coding\nBoth move fast, but AI prototyping is about testing the product idea.\n\nHuman-in-the-loop\nEven a fast prototype still needs people to approve the big calls.", "relations": { "copilot": { "label": "often uses …", "note": "Copilot is often the hands-on helper for a first prototype." }, "agentic-coding": { "label": "can grow into …", "note": "Agentic coding goes further, with AI changing more code on its own." }, "vibe-coding": { "label": "gets confused with …", "note": "AI prototyping tests the idea, not just the fun of coding fast." }, "human-in-the-loop": { "label": "still needs …", "note": "A fast prototype still needs a person to judge it." } } }, "zh": { "fullName": "AI 加速原型开发", "factExplain": "用 AI 更快产出产品原型的工作方式。", "humanExplain": "AI 原型设计像先点外卖小份试吃，别一上来就承包整桌年夜饭。\n\n它适合做产品试水、流程演示，把想法更快变成可讨论的东西。", "humanExplainDisplay": "AI 原型设计像==先点外卖小份试吃==，\n别一上来就==承包整桌年夜饭==。\n\n它适合做产品试水、流程演示，\n把想法更快变成可讨论的东西。", "relationsNarrative": "Copilot\nAI 加速原型开发常借助 Copilot 快速产出初版。\n\nAgentic coding\n它再往前一步，就会走向更自主的 Agentic coding。\n\nVibe-coding\n两者都追求快，但它更强调验证产品想法。\n\nHuman-in-the-loop\n原型做得再快，关键决策和验收仍要人把关。", "relations": { "copilot": { "label": "常借助…完成", "note": "Copilot 常是原型开发的直接助手。" }, "agentic-coding": { "label": "可升级为…", "note": "从补写原型到自主改代码更进一步。" }, "vibe-coding": { "label": "常被误当成…", "note": "它重在验证想法，不只图写得爽。" }, "human-in-the-loop": { "label": "仍需要…把关", "note": "原型快不等于可上线，仍要人判断。" } } } } }, { "id": "ai-accommodation", "name": "AI Accommodation", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "ai-academic-integrity" }, { "to": "ai-school-ban" }, { "to": "ai-usage-gap" }, { "to": "ai-literacy-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Accommodation", "factExplain": "An arrangement that adjusts AI rules for people with specific needs.", "humanExplain": "AI accommodation is like giving left-handed scissors to a left-handed kid. It is not cheating. It just makes craft time fair.\n\nYou see it at school and work. It also supports accessibility and keeps AI use within the rules.", "humanExplainDisplay": "AI accommodation is like giving ==left-handed scissors==\nto a left-handed kid.\nIt is ==not cheating==.\nIt just makes craft time fair.\n\nYou see it at school and work.\nIt also supports accessibility\nand keeps AI use within the rules.", "relationsNarrative": "AI Academic Integrity\nAI accommodation helps schools separate fair support from cheating.\n\nAI School Ban\nAI accommodation is more careful than banning AI for everyone.\n\nAI usage gap\nAI accommodation helps disadvantaged people use AI more fairly.\n\nAI Literacy\nAI accommodation works best when people can use AI and explain their use.", "relations": { "ai-academic-integrity": { "label": "sets boundaries for …", "note": "Helpful support and cheating need a clear line." }, "ai-school-ban": { "label": "offers an alternative to …", "note": "An accommodation is more careful than a full ban." }, "ai-usage-gap": { "label": "narrows …", "note": "Clear support helps more people get fair AI access." }, "ai-literacy-ai": { "label": "depends on …", "note": "People need to know how to use AI and explain it." } } }, "zh": { "fullName": "AI 便利安排", "factExplain": "为特定人群调整 AI 使用规则的安排。", "humanExplain": "AI 便利安排像考场给近视娃配眼镜：不是放水，是先让人看清题。\n\n用于学校、职场和无障碍场景，帮人合规用 AI。", "humanExplainDisplay": "AI 便利安排像考场\n给==近视娃配眼镜==：\n==不是放水==，\n是先让人看清题。\n\n用于学校、职场和无障碍场景，\n帮人合规用 AI。", "relationsNarrative": "AI Academic Integrity\n它帮学校区分合理辅助和作弊边界。\n\nAI School Ban\n它是比一刀切禁用更细的治理办法。\n\nAI Usage Gap\n它能让弱势群体更公平地使用 AI。\n\nAI Literacy\n会用也会说明，便利安排才不易越界。", "relations": { "ai-academic-integrity": { "label": "划清…边界", "note": "合理辅助和作弊要分清。" }, "ai-school-ban": { "label": "替代一刀切…", "note": "便利安排比全禁更细。" }, "ai-usage-gap": { "label": "缩小…", "note": "有规则支持，弱势群体更能用上。" }, "ai-literacy-ai": { "label": "依赖…", "note": "会用、会说明，才不容易越界。" } } } } }, { "id": "ai-adoption-curve", "name": "Adoption Curve", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "ai-anxiety" }, { "to": "ai-native-organization" }, { "to": "automation-job" } ], "track": "today", "i18n": { "en": { "fullName": "Adoption Curve", "factExplain": "A pattern that shows how AI spreads from early users to everyday use.", "humanExplain": "A new app starts with tech kids. Then your uncle uses it. That is the adoption curve.\n\nIt shows if AI is still a toy for a few people. Or already part of normal work and life.", "humanExplainDisplay": "A new app starts\nwith ==tech kids==.\nThen ==your uncle uses it==.\nThat is the adoption curve.\n\nIt shows if AI is still\na toy for a few people.\nOr already part of normal work and life.", "relationsNarrative": "AI-anxiety\nWhen AI spreads fast, more people fear being left behind.\n\nAI-native organization\nWhen adoption gets deep, organizations start rebuilding work around AI.\n\nAutomation-job\nThe steepness of the adoption curve affects how fast jobs change.", "relations": { "ai-anxiety": { "label": "helps explain …", "note": "When AI spreads fast, more people fear being left behind." }, "ai-native-organization": { "label": "pushes …", "note": "Deeper adoption makes organizations rebuild work around AI." }, "automation-job": { "label": "sets the pace for …", "note": "Faster adoption changes jobs faster." } } }, "zh": { "fullName": "AI 采用曲线", "factExplain": "描述 AI 技术从少数尝鲜到大众普及的扩散过程。", "humanExplain": "采用曲线像小区团购：有人秒下单，有人等全群晒图才跟。\n\n它帮你判断谁先试、谁观望，适合看产品扩散、组织落地。", "humanExplainDisplay": "采用曲线像==小区团购==：\n有人秒下单，\n有人等==全群晒图==才跟。\n\n它帮你判断谁先试、谁观望，\n适合看产品扩散、组织落地。", "relationsNarrative": "AI-anxiety\nAI 普及越快，普通人越容易产生被落下的焦虑。\n\nAI-native organization\n采用进入深水区后，组织会开始按 AI 重构流程。\n\nAutomation-job\n采用曲线的陡峭程度，会影响岗位变化节奏。", "relations": { "ai-anxiety": { "label": "解释…来源", "note": "普及加速时，普通人更易产生焦虑。" }, "ai-native-organization": { "label": "推动…出现", "note": "采用越深，组织越可能重构流程。" }, "automation-job": { "label": "影响…节奏", "note": "采用快慢会改变岗位受冲击速度。" } } } } }, { "id": "ai-ai-code-review", "name": "AI Code Review", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "copilot" }, { "to": "cursor" }, { "to": "human-in-the-loop" }, { "to": "ai-vulnerability-discovery-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Code Review", "factExplain": "AI that checks code for problems and suggests fixes.", "humanExplain": "It is like a picky friend checking your LEGO bridge. They poke wobbly bits before the cat does.\n\nIt checks code in pull requests and security checks, then suggests fixes. It saves time, but a human still approves the merge.", "humanExplainDisplay": "It is like a ==picky friend==\nchecking your LEGO bridge.\nThey poke ==wobbly bits==\nbefore the cat does.\n\nIt checks code in pull requests and security checks,\nthen suggests fixes.\nIt saves time,\nbut a human still approves the merge.", "relationsNarrative": "Copilot\nCopilot helps write code. AI Code Review checks it and suggests fixes.\n\nCursor\nCursor can build AI Code Review into the coding flow.\n\nHuman-in-the-loop\nAI Code Review can flag issues first, but a human decides what to merge.\n\nAI Vulnerability Discovery\nAI Code Review can find clues that help Vuln Discovery.", "relations": { "copilot": { "label": "checks after … writes", "note": "Copilot helps write code. AI Code Review helps find mistakes." }, "cursor": { "label": "is often built into …", "note": "Many smart editors include AI code review." }, "human-in-the-loop": { "label": "needs … to decide", "note": "The AI can advise, but people own the merge." }, "ai-vulnerability-discovery-ai": { "label": "can spot clues for …", "note": "Code review can catch some security bugs and risky code." } } }, "zh": { "fullName": "AI 代码审查", "factExplain": "用 AI 自动检查代码问题并给出修改建议。", "humanExplain": "像收房前的验房师，敲墙、拧水管，哪儿空鼓哪儿漏水先红笔圈出来；修不修、收不收，还是你拍板。\n\n常用于 PR 审查和安全检查，能提速提质，但最终合并仍要人拍板。", "humanExplainDisplay": "像收房前的==验房师==，\n敲墙、拧水管，\n哪儿空鼓哪儿漏水==先圈出来==；\n修不修、收不收，还是你拍板。\n\n常用于 PR 审查和安全检查，\n能提速提质，\n但最终合并仍要人拍板。", "relationsNarrative": "Copilot\nCopilot 偏重写代码，它偏重检查和提建议。\n\nCursor\n很多智能编辑器会把它直接做进开发流程里。\n\nHuman-in-the-loop\n它能先筛问题，但最终是否采纳仍要人判断。\n\nAI Vulnerability Discovery\n代码审查常会顺带发现漏洞和危险写法。", "relations": { "copilot": { "label": "补上…后半程", "note": "前者帮你写，后者帮你挑错。" }, "cursor": { "label": "常被集成进…", "note": "很多智能编辑器内置这类审查能力。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "建议能参考，但合并责任仍在人。" }, "ai-vulnerability-discovery-ai": { "label": "可发现…线索", "note": "它能顺手揪出部分安全漏洞。" } } } } }, { "id": "ai-ai-fatigue", "name": "AI fatigue", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "ai-anxiety" }, { "to": "ai-adoption-curve" }, { "to": "permission-fatigue" } ], "track": "today", "i18n": { "en": { "fullName": "AI fatigue", "factExplain": "Tiredness and pushback people feel after too much exposure to AI tools.", "humanExplain": "AI fatigue is like every app on your phone yelling, “Now with AI!” By Friday, even your weather app wants to write a poem.\n\nIt shows up when many AI tools arrive at once. People stop trying them, and may turn off useful features.", "humanExplainDisplay": "AI fatigue is like every app on your phone yelling,\n==“Now with AI!”==\nBy Friday,\neven your ==weather app wants to write a poem==.\n\nIt shows up when many AI tools arrive at once.\nPeople stop trying them,\nand may turn off useful features.", "relationsNarrative": "AI-anxiety\nFear of being replaced can make AI fatigue worse.\n\nAdoption Curve\nAI fatigue can turn early interest into a slow rollout.\n\nPermission fatigue\nBoth start with too many interruptions, so people click no.", "relations": { "ai-anxiety": { "label": "is often worsened by …", "note": "More fear makes new tools feel annoying faster." }, "ai-adoption-curve": { "label": "slows …", "note": "Fatigue cools the urge to try AI at work." }, "permission-fatigue": { "label": "feels like …", "note": "Both come from too many pings and too many asks." } } }, "zh": { "fullName": "AI 疲劳", "factExplain": "对 AI 工具的过度接触后产生的厌倦与抗拒。", "humanExplain": "像相亲角一周给你安排八场，刚开始还会看看条件，后来别人一提“优质对象”你就想撤。\n\n多见于工具密集上线时，会拖慢采纳，也让用户顺手关掉本来有用的功能。", "humanExplainDisplay": "像相亲角一周给你安排==八场==，\n刚开始还会看看条件，\n后来别人一提“==优质对象==”\n你就想撤。\n\n多见于工具密集上线时，\n会拖慢采纳，\n也让用户顺手关掉本来有用的功能。", "relationsNarrative": "AI-anxiety\n对被替代或掉队的担心，常会加重这种厌倦感。\n\nAdoption Curve\n它会让个人和组织从尝鲜转向冷淡，拖慢采纳。\n\nPermission fatigue\n两者都源于高频打扰，最后让人下意识点拒绝。", "relations": { "ai-anxiety": { "label": "常被…放大", "note": "焦虑越强，人越容易对新工具生厌。" }, "ai-adoption-curve": { "label": "会拖慢…", "note": "疲劳会让组织采纳热情明显降温。" }, "permission-fatigue": { "label": "和…很像", "note": "都是被反复打扰后产生麻木抗拒。" } } } } }, { "id": "ai-ai-note-taking", "name": "AI Notes", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "rag" }, { "to": "agent-memory" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "AI Notes", "factExplain": "A note app using AI to record, sort, and sum up information.", "humanExplain": "AI Notes is the kid in class with twelve highlighters. You blink, and it has the key points and homework list.\n\nIt records messy talk. Then it turns it into clean notes. You see it after meetings, classes, and interviews. It saves time. Still check the important parts yourself.", "humanExplainDisplay": "AI Notes is the ==kid in class==\nwith twelve highlighters.\nYou blink,\nand it has the ==key points==\nand the homework list.\n\nIt records messy talk.\nThen it turns it into clean notes.\nYou see it after meetings, classes, and interviews.\nIt saves time.\nStill check the important parts yourself.", "relationsNarrative": "STT\nMany AI Notes tools turn speech into text first, then make a summary.\n\nRAG\nAI Notes can use RAG to search old notes for past details.\n\nMemory\nNotes can work as outside long-term memory for a system.\n\nData-privacy\nMeetings, classes, and private notes can create privacy risks.", "relations": { "speech-to-text": { "label": "uses … for transcripts", "note": "STT turns speech into text before AI Notes cleans it up." }, "rag": { "label": "finds old notes with …", "note": "RAG helps search past notes for the right answer." }, "agent-memory": { "label": "acts like outside …", "note": "Notes can act as long-term memory for a system." }, "data-privacy": { "label": "raises … risks", "note": "Meeting and personal notes can hold sensitive information." } } }, "zh": { "fullName": "AI 笔记", "factExplain": "用 AI 辅助记录、整理和提炼信息的笔记应用。", "humanExplain": "上课你还在追老师板书，它已经像前排课代表，记重点、划考点、顺手把作业列好了。\n\n常用于会议纪要、课程整理和采访转写，省时间，但关键信息仍要自己复核。", "humanExplainDisplay": "上课你还在追老师板书，\n它已经像==前排课代表==，\n记重点、划考点，\n顺手把==作业列好==了。\n\n常用于会议纪要、课程整理和采访转写，\n省时间，但关键信息仍要自己复核。", "relationsNarrative": "Speech-to-text\n很多 AI 笔记先把语音转成文字，再做摘要和整理。\n\nRAG\n当它要回答“以前记过什么”时，常靠检索旧笔记。\n\nMemory\n笔记常被当成系统的外部长期记忆来调用。\n\nData-privacy\n会议、课程和私人记录，都会带来隐私风险。", "relations": { "speech-to-text": { "label": "常用…做转写", "note": "先把语音变文字，才好继续整理。" }, "rag": { "label": "可配合…找旧内容", "note": "翻历史笔记时，常靠检索增强回答。" }, "agent-memory": { "label": "像…的外部脑", "note": "笔记能给系统提供长期可查的记忆。" }, "data-privacy": { "label": "会碰到…问题", "note": "会议和私密记录常涉及敏感信息。" } } } } }, { "id": "ai-anti-cheat", "name": "AI Anti-Cheat", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "ai-proctoring" }, { "to": "ai-detector" }, { "to": "captcha" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "AI Anti-Cheat", "factExplain": "Tools and rules that spot and block cheating with AI help.", "humanExplain": "AI Anti-Cheat is a hall monitor with Wi-Fi paranoia. It checks for the robot hiding in your hoodie.\n\nYou meet it in online exams. You also see it in hiring tests and contest sites. It blocks stand-ins, scripts, and AI helpers.", "humanExplainDisplay": "AI Anti-Cheat is a ==hall monitor==\nwith Wi-Fi paranoia.\nIt checks for the ==robot hiding in your hoodie==.\n\nYou meet it in online exams.\nYou also see it in hiring tests and contest sites.\nIt blocks stand-ins, scripts, and AI helpers.", "relationsNarrative": "AI Proctoring\nAI Anti-Cheat pairs with AI Proctoring: one watches the person, the other the trick.\n\nAI detector\nAI Anti-Cheat can use an AI detector to spot AI-made answers or behavior.\n\nCAPTCHA\nAI Anti-Cheat and CAPTCHA both check if a real person did the work alone.\n\nAI-regulation\nAI-regulation affects how strict AI Anti-Cheat can be.", "relations": { "ai-proctoring": { "label": "pairs with …", "note": "AI Proctoring watches the person. Anti-Cheat watches the trick." }, "ai-detector": { "label": "spots AI traces with …", "note": "An AI detector can flag answers that look AI-made." }, "captcha": { "label": "continues the idea of …", "note": "Both ask if a real person did the work alone." }, "ai-regulation": { "label": "is shaped by …", "note": "School and platform rules decide how strict the checks can be." } } }, "zh": { "fullName": "AI 反作弊", "factExplain": "用来识别和拦截 AI 辅助作弊的检测与防范机制。", "humanExplain": "现在防作弊不像只看你像不像本人，更像直播打假：人是你出镜，也得查台词、提词器和后台代操盘。\n\n常用于线上考试、招聘测评和竞赛平台，拦截代答、脚本和 AI 帮手。", "humanExplainDisplay": "现在防作弊不像只看你\n像不像本人，更像直播\n==打假==：人是你出镜，也得查\n台词、提词器和后台\n==代操盘==。\n\n常用于线上考试、招聘测评\n和竞赛平台，拦截代答、\n脚本和 AI 帮手。", "relationsNarrative": "AI Proctoring\n常和 AI 监考一起用：一个盯人，一个盯作弊手段。\n\nAI Detector\n它常借助检测器判断答案或行为是否有 AI 痕迹。\n\nCAPTCHA\n两者都在分辨到底是真人独立完成，还是机器掺了一脚。\n\nAI-regulation\n学校、平台和监管规则，会决定它能管多严、查多深。", "relations": { "ai-proctoring": { "label": "常与…配套", "note": "监考负责盯人，它更偏盯作弊手段。" }, "ai-detector": { "label": "常借助…识别", "note": "可用检测器判断内容是否像 AI 生成。" }, "captcha": { "label": "延续…思路", "note": "都在区分真人操作和机器参与。" }, "ai-regulation": { "label": "受…推动", "note": "学校和平台规则会影响部署强度。" } } } } }, { "id": "ai-anxiety", "name": "AI-anxiety", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T10:55:00Z", "relations": [ { "to": "automation-job" }, { "to": "agi" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "AI Anxiety", "factExplain": "Unease people feel as AI changes skills, jobs, and society.", "humanExplain": "AI anxiety feels like seeing a robot join your group project. It never sleeps, and it might work for pizza money.\n\nIt often comes from work stress, school stress, and feeling replaced. Knowing AI’s limits helps more than panic.", "humanExplainDisplay": "AI anxiety feels like seeing a ==robot join your group project==.\nIt ==never sleeps==,\nand it might work for pizza money.\n\nIt often comes from work stress,\nschool stress,\nand feeling replaced.\nKnowing AI’s limits helps more than panic.", "relationsNarrative": "Automation-job\nAutomation-job is a direct job-level source of AI anxiety.\n\nAGI\nAGI adds fear about AI getting out of control someday.\n\nAI-regulation\nAI-regulation can ease AI anxiety with rules and safety nets.", "relations": { "automation-job": { "label": "comes from …", "note": "Job automation is a direct work source of AI anxiety." }, "agi": { "label": "worries about …", "note": "AGI adds fear about AI getting out of control someday." }, "ai-regulation": { "label": "can be eased by …", "note": "AI regulation can give society a safety net." } } }, "zh": { "fullName": "AI 焦虑", "factExplain": "人们因 AI 发展带来的能力、职业和社会变化产生的不安。", "humanExplain": "AI 焦虑像同事都在用外挂赶工，你还在家庭群问“这按钮咋点”。\n\n它会影响学习、求职和管理决策，适合用小步试用，慢慢拆掉恐慌。", "humanExplainDisplay": "AI 焦虑像同事都在==用外挂赶工==，\n你还在家庭群问“==这按钮咋点==”。\n\n它会影响学习、求职和管理决策，\n适合用小步试用，\n慢慢拆掉恐慌。", "relationsNarrative": "Automation-job\nAutomation-job 是 AI-anxiety 在就业层面的直接来源。\n\nAGI\nAGI 放大了 AI-anxiety 中关于长期失控的担忧。\n\nAI-regulation\nAI-regulation 为 AI-anxiety 提供制度层面的缓冲。", "relations": { "automation-job": { "label": "源于…" }, "agi": { "label": "源于对…的担忧" }, "ai-regulation": { "label": "可被…缓解" } } } } }, { "id": "ai-app-builder", "name": "AI App Builder", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "api" }, { "to": "ai-accelerated-prototyping" }, { "to": "personal-ai-apps" } ], "track": "today", "i18n": { "en": { "fullName": "AI App Builder", "factExplain": "A tool for quickly making AI apps with prompts or little code.", "humanExplain": "An AI App Builder is like a boxed cake mix for software. You say what you want, and it gives you something you can click.\n\nIt helps with office tools, support bots, and quick demos. Big messy systems still need engineers to finish the job.", "humanExplainDisplay": "An AI App Builder is like a ==boxed cake mix== for software.\nYou say what you want,\nand it gives you ==something you can click==.\n\nIt helps with office tools,\nsupport bots,\nand quick demos.\nBig messy systems still need engineers\nto finish the job.", "relationsNarrative": "LLM\nAn AI App Builder often uses an LLM to write and understand things.\n\nAPI\nAPIs let it connect to models, data, and business tools.\n\nAI Prototyping\nIt makes AI prototypes feel more like building with blocks.\n\nPersonal AI apps\nIt gives regular people a way to build small AI tools.", "relations": { "llm": { "label": "uses … for generation", "note": "Most app builders use an LLM as the main engine." }, "api": { "label": "connects services through …", "note": "APIs connect models, data, and business tools." }, "ai-accelerated-prototyping": { "label": "speeds up …", "note": "It turns ideas into testable prototypes faster." }, "personal-ai-apps": { "label": "creates …", "note": "Many small personal AI tools can be built with it." } } }, "zh": { "fullName": "AI 应用构建器", "factExplain": "用低代码或提示词快速生成 AI 应用的工具。", "humanExplain": "AI 应用构建器像点单即出的快厨：你说要什么应用，它当场端出能点能用的成品。\n\n适合内部工具、客服和原型，快成型，复杂系统仍要工程收尾。", "humanExplainDisplay": "AI 应用构建器像点单即出的快厨：\n你说要什么应用，\n它当场端出==能点能用的成品==。\n\n适合内部工具、客服\n和原型，\n快成型，复杂系统仍要工程收尾。", "relationsNarrative": "LLM\n应用构建器通常把 LLM 当生成和理解核心。\n\nAPI\nAPI 让它接入模型、数据和业务系统。\n\nAI Prototyping\n它把 AI 原型从“写代码”变成“搭积木”。\n\nPersonal AI apps\n它是普通人做个人 AI 小工具的入口。", "relations": { "llm": { "label": "调用…生成能力", "note": "多数构建器把 LLM 当核心引擎。" }, "api": { "label": "通过…接服务", "note": "API 负责连模型、数据和业务系统。" }, "ai-accelerated-prototyping": { "label": "加速…", "note": "它能把想法更快变成可试原型。" }, "personal-ai-apps": { "label": "产出…", "note": "很多个人小工具可由它快速搭出。" } } } } }, { "id": "ai-artifacts", "name": "AI Artifacts", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "chatgpt" }, { "to": "agentic-coding" }, { "to": "structured-output" }, { "to": "copilot" } ], "track": "today", "i18n": { "en": { "fullName": "AI Artifacts", "factExplain": "AI-made finished work you can save, use, or hand over.", "humanExplain": "A chatbot is a classmate giving tips. An AI Artifact is the finished poster for the science fair.\n\nYou meet them as code files and documents. You also see sheets and web mockups. You can edit them. You can share them. You can ship them.", "humanExplainDisplay": "A chatbot is a classmate giving tips.\nAn AI Artifact is the ==finished poster==\nfor the ==science fair==.\n\nYou meet them as code files and documents.\nYou also see sheets and web mockups.\nYou can edit them.\nYou can share them.\nYou can ship them.", "relationsNarrative": "ChatGPT\nMany users first meet AI Artifacts inside a chat product.\n\nAgentic coding\nAgentic coding often produces files you can keep editing.\n\nStructured output\nStructured output makes artifacts steadier and easier for systems to use.\n\nCopilot\nCopilot tools help teams make and deliver these artifacts.", "relations": { "chatgpt": { "label": "often comes from …", "note": "Many people first see AI Artifacts inside a chat window." }, "agentic-coding": { "label": "is delivered by …", "note": "Coding agents often produce finished files you can keep editing." }, "structured-output": { "label": "can be shaped by …", "note": "Structured output helps artifacts fit into other systems." }, "copilot": { "label": "gets made inside …", "note": "Copilot tools are common places to write and revise artifacts." } } }, "zh": { "fullName": "AI 生成成品／可交付物", "factExplain": "由 AI 生成并可直接保存、使用或交付的内容成品。", "humanExplain": "不是陪你唠两句的客服，而是打印店老板把活都装订好了：拿走就能交。\n\n常见于代码、文档、表格和网页原型，适合继续修改、协作和落地。", "humanExplainDisplay": "不是陪你唠两句的客服，\n而是打印店老板\n把活都==装订好了==：\n==拿走就能交==。\n\n常见于代码、文档、表格\n和网页原型，\n适合继续修改、协作\n和落地。", "relationsNarrative": "ChatGPT\n很多用户最早在聊天产品里接触到这类 AI 成品。\n\nAgentic coding\n写代码型智能体常直接产出可继续修改的成品。\n\nStructured output\n结构化输出能让生成成品更稳定、更方便接系统。\n\nCopilot\n这类成品常在协作工具里生成、编辑和交付。", "relations": { "chatgpt": { "label": "常由…产出", "note": "很多人第一次接触它就在聊天窗口里。" }, "agentic-coding": { "label": "是…的交付物", "note": "写代码型智能体常直接产出这类成品。" }, "structured-output": { "label": "可借…规范", "note": "结构化约束能让成品更好接系统。" }, "copilot": { "label": "在…里落地", "note": "编写与修改成品常发生在协作工具中。" } } } } }, { "id": "ai-assisted-research", "name": "AI-assisted Research", "layer": "L4", "era": "2023", "publishedAt": "2026-05-31T00:57:33.088Z", "relations": [ { "to": "copilot" }, { "to": "rag" }, { "to": "hallucination" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI-assisted Research", "factExplain": "Using AI to find, sort, analyze, and draft research work.", "humanExplain": "It is a super-fast library kid with a backpack full of sticky notes. It finds books fast, then sometimes invents one with a very serious face.\n\nUse it to search papers and organize notes. It can draft early text, but a human must check conclusions.", "humanExplainDisplay": "It is a ==super-fast library kid==\nwith a backpack full of sticky notes.\nIt finds books fast,\nthen sometimes ==invents one==\nwith a very serious face.\n\nUse it to search papers\nand organize notes.\nIt can draft early text,\nbut a human must check conclusions.", "relationsNarrative": "Copilot\nAI-assisted research often works as a Copilot inside search, writing, and office tools.\n\nRAG\nIt often uses RAG to find papers, web pages, or internal files first.\n\nHallucination\nAI-assisted research saves time, but it may invent sources or twist meanings.\n\nHuman-in-the-loop\nDo not trust AI alone for research conclusions. A human must check the key steps.", "relations": { "copilot": { "label": "often works as …", "note": "Many research tools appear as a Copilot in search, writing, and office work." }, "rag": { "label": "looks things up with …", "note": "RAG helps it fetch outside sources before it summarizes." }, "hallucination": { "label": "can fall into …", "note": "AI may invent sources or misread results." }, "human-in-the-loop": { "label": "needs … to check it", "note": "A researcher must still make the key calls." } } }, "zh": { "fullName": "AI 辅助研究", "factExplain": "用 AI 协助检索、整理、分析与写作的研究方式。", "humanExplain": "AI 辅助研究像给科研打工人配了个嘴快助理，翻资料比抢红包还积极。\n\n它常用于文献梳理、实验设计和报告初稿，但结论必须人工复核。", "humanExplainDisplay": "AI 辅助研究像给科研打工人\n配了个==嘴快助理==，\n翻资料==比抢红包还积极==。\n\n它常用于文献梳理、\n实验设计和报告初稿，\n但结论必须人工复核。", "relationsNarrative": "Copilot\nAI-assisted research 常以 Copilot 形式嵌入搜索、写作和办公流程。\n\nRAG\n它常借助 RAG 检索论文、网页或内部资料再做整理。\n\nHallucination\nAI 辅助研究提效明显，但也可能捏造来源或曲解内容。\n\nHuman-in-the-loop\n研究结论不能只信 AI，关键环节仍需要人复核。", "relations": { "copilot": { "label": "常以…落地", "note": "很多研究场景以 Copilot 形态出现。" }, "rag": { "label": "常靠…找资料", "note": "RAG 帮它先找外部材料再总结。" }, "hallucination": { "label": "容易出现…", "note": "AI 可能编造文献或误读结论。" }, "human-in-the-loop": { "label": "需要…复核", "note": "关键判断仍需研究者亲自把关。" } } } } }, { "id": "ai-audio-generation", "name": "Audio Generation", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "tts" }, { "to": "voice-cloning" }, { "to": "deepfake" }, { "to": "multimodal" } ], "track": "today", "i18n": { "en": { "fullName": "AI Audio Generation", "factExplain": "AI that creates speech, music, or sound effects.", "humanExplain": "It is like the school play sound kid with too many buttons. You ask for a robot voice, and somehow you also get thunder.\n\nYou meet it in voice-overs and music tools. It can also fake a real person's voice.", "humanExplainDisplay": "It is like the ==school play sound kid==\nwith ==too many buttons==.\nYou ask for a robot voice,\nand somehow you also get thunder.\n\nYou meet it in voice-overs\nand music tools.\nIt can also fake\na real person's voice.", "relationsNarrative": "TTS\nTTS is one of the most common uses of audio generation.\n\nVoice cloning\nVoice cloning goes further by copying a specific voice.\n\nDeepfake\nRealistic audio generation can make fake recordings more dangerous.\n\nMultimodal AI\nAudio generation handles sound, so it often sits inside Multimodal AI.", "relations": { "tts": { "label": "includes …", "note": "TTS is the most common branch of audio generation." }, "voice-cloning": { "label": "can extend into …", "note": "Voice cloning imitates a specific person's voice." }, "deepfake": { "label": "raises … risk", "note": "Realistic audio can make fake recordings easier." }, "multimodal": { "label": "is part of …", "note": "Audio generation works with the sound part of AI." } } }, "zh": { "fullName": "AI 音频生成", "factExplain": "用模型生成语音、音乐或音效的技术。", "humanExplain": "像街头那种一人乐队，背上鼓、嘴上口琴、脚下镲，你点什么风格，当场就配出一整套声音。\n\n常用于配音和音乐制作，也可能被拿来伪造声音。", "humanExplainDisplay": "像街头那种==一人乐队==，\n背上鼓、嘴上口琴、脚下镲，\n你点什么风格，\n当场就==配出一整套声音==。\n\n常用于配音\n和音乐制作，\n也可能被拿来伪造声音。", "relationsNarrative": "TTS\n语音合成是音频生成里最常见、最成熟的一类应用。\n\nVoice cloning\n声音克隆是音频生成的进一步能力，重点在模仿特定声线。\n\nDeepfake\n逼真音频生成会放大伪造录音和冒充他人的风险。\n\nMultimodal\n音频生成处理声音模态，常是多模态系统的一部分。", "relations": { "tts": { "label": "包含…能力", "note": "语音合成是它最常见分支。" }, "voice-cloning": { "label": "延伸到…", "note": "可进一步模仿特定人的声音。" }, "deepfake": { "label": "催生…风险", "note": "生成逼真音频也会带来伪造问题。" }, "multimodal": { "label": "属于…应用", "note": "它处理的是声音这一模态。" } } } } }, { "id": "ai-bias", "name": "AI-bias", "layer": "L6", "era": "2018", "publishedAt": "2026-05-23T11:25:00Z", "relations": [ { "to": "data-privacy" }, { "to": "alignment" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "Algorithmic Bias", "factExplain": "An unfair tilt in AI caused by its data, design, or real-world use.", "humanExplain": "AI bias is like a school lunch monitor with a whistle. It says it is fair, then gives the best pizza to its favorite table.\n\nIt can shape hiring, loans, your feed, and app post checks. So do not treat “the computer said so” as fair.", "humanExplainDisplay": "AI bias is like a ==school lunch monitor==\nwith a whistle.\nIt says it is fair,\nthen gives the ==best pizza==\nto its favorite table.\n\nIt can shape hiring, loans,\nyour feed, and app post checks.\nSo do not treat “the computer said so”\nas fair.", "relationsNarrative": "Data-privacy\nData-privacy shapes the data AI can use, so it can shape AI bias.\n\nAlignment\nAlignment tries to reduce AI bias in model behavior.\n\nAI-regulation\nAI-regulation asks systems to find, report, and lower AI bias.", "relations": { "data-privacy": { "label": "is affected by …", "note": "Data-privacy shapes training data, so it can shape bias." }, "alignment": { "label": "is reduced by …", "note": "Alignment tries to make models act with less unfair bias." }, "ai-regulation": { "label": "is watched by …", "note": "AI-regulation asks teams to find, report, and lower bias." } } }, "zh": { "fullName": "算法偏见", "factExplain": "AI 因数据、设计或使用环境产生不公平倾向的现象。", "humanExplain": "AI 偏见就像相亲阿姨只看户口本，资料里偏啥，它就跟着偏啥。\n\n它会影响招聘、信贷、内容推荐，需要评测和治理来兜底。", "humanExplainDisplay": "AI 偏见就像==相亲阿姨只看户口本==，\n资料里偏啥，\n它就跟着偏啥。\n\n它会影响招聘、信贷、内容推荐，\n需要评测和治理来兜底。", "relationsNarrative": "Data-privacy\nData-privacy 会影响训练数据边界，从而影响 AI-bias。\n\nAlignment\nAlignment 试图减少 AI-bias 在模型行为中的表现。\n\nAI-regulation\nAI-regulation 要求系统识别、披露并降低 AI-bias。", "relations": { "data-privacy": { "label": "受…影响" }, "alignment": { "label": "由…改善" }, "ai-regulation": { "label": "被…关注" } } } } }, { "id": "ai-biosecurity", "name": "AI biosecurity", "layer": "L6", "era": "2023", "publishedAt": "2026-05-30T03:10:23.230Z", "relations": [ { "to": "ai-regulation" }, { "to": "frontier-model" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "AI Biosecurity", "factExplain": "Safety work that stops AI from helping with dangerous biology misuse.", "humanExplain": "AI biosecurity is the lock on the science-room cabinet. Helpful lab tips come out, but the creepy how-to card stays inside.\n\nIt sets limits for biology answers from AI. It matters in model tests, releases, and rules.", "humanExplainDisplay": "AI biosecurity is the ==lock==\non the science-room cabinet.\nHelpful lab tips come out,\nbut the ==creepy how-to card== stays inside.\n\nIt sets limits for biology answers from AI.\nIt matters in model tests, releases, and rules.", "relationsNarrative": "AI-regulation\nAI biosecurity pushes regulation to set red lines for high-risk biology skills.\n\nFrontier model\nFrontier models are stronger, so they need special biosecurity tests.\n\nAlignment\nAI biosecurity is part of Alignment because the goal is safer model behavior.", "relations": { "ai-regulation": { "label": "pushes … to set red lines", "note": "High-risk biology skills often need rules and review." }, "frontier-model": { "label": "checks … for bio risk", "note": "Stronger models need tougher biosecurity tests." }, "alignment": { "label": "is part of …", "note": "It helps stop models from aiding dangerous uses." } } }, "zh": { "fullName": "AI 生物安全", "factExplain": "研究并防范 AI 被用于生物风险的安全议题。", "humanExplain": "AI 生物安全像管住实验室钥匙，别让抢答学霸帮人乱配危险药方。\n\n它用于模型发布、实验室审查和研究助手，防误用，也保正常科研。", "humanExplainDisplay": "AI 生物安全像==管住实验室钥匙==，\n别让==抢答学霸==帮人乱配危险药方。\n\n它用于模型发布、\n实验室审查和研究助手，\n防误用，也保正常科研。", "relationsNarrative": "AI-regulation\nAI 生物安全常推动监管为高风险能力设置红线和审查要求。\n\nFrontier model\n前沿模型能力更强，也更需要接受生物安全方向的专门评估。\n\nAlignment\nAI 生物安全是对齐问题的一部分，目标是让模型别去帮忙做危险的事。", "relations": { "ai-regulation": { "label": "推动…设红线", "note": "高风险生物能力常需监管约束。" }, "frontier-model": { "label": "重点评估…风险", "note": "能力越强的模型越需生物安全测试。" }, "alignment": { "label": "属于…关注面", "note": "让模型别助长危险用途。" } } } } }, { "id": "ai-bot-traffic", "name": "AI bot traffic", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "api" }, { "to": "agent" }, { "to": "captcha" }, { "to": "ai-usage-cap" } ], "track": "today", "i18n": { "en": { "fullName": "AI bot traffic", "factExplain": "Website visits made by automated programs instead of people.", "humanExplain": "It is like robot shoppers grabbing every concert ticket. They never go to the show, but your cart still says sold out.\n\nYou meet it in crawlers and API calls. When it runs wild, it eats bandwidth and raises the site's bill.", "humanExplainDisplay": "It is like ==robot shoppers==\ngrabbing every concert ticket.\nThey never go to the show,\nbut your cart still says ==sold out==.\n\nYou meet it in crawlers and API calls.\nWhen it runs wild,\nit eats bandwidth and raises the site's bill.", "relationsNarrative": "API\nMany bots skip web pages and hit the API again and again.\n\nAgent\nAgents can run tasks in bulk, so they can create bot traffic.\n\nCAPTCHA\nCAPTCHA helps block bot visits and let real people through.\n\nUsage cap\nWhen bots flood a service, a usage cap helps slow them down.", "relations": { "api": { "label": "hammers … endpoints", "note": "Many bots skip pages and hit APIs again and again." }, "agent": { "label": "often comes from …", "note": "Agents can run many machine tasks at once." }, "captcha": { "label": "is often blocked by …", "note": "CAPTCHA helps tell people from bots." }, "ai-usage-cap": { "label": "pushes platforms toward …", "note": "Usage caps slow down heavy bot floods." } } }, "zh": { "fullName": "AI 机器人流量", "factExplain": "由自动化程序发起的网络访问流量。", "humanExplain": "它像黄牛抢演唱会票：自己根本不看演出，却把票全扫空，真粉丝只能干瞪眼。\n\n它常见于爬虫和自动调用，失控时会挤占带宽并抬高网站成本。", "humanExplainDisplay": "它像==黄牛==抢演唱会票：\n自己根本不看演出，\n却把票全扫空，\n真粉丝只能==干瞪眼==。\n\n它常见于爬虫和自动调用，\n失控时会挤占带宽\n并抬高网站成本。", "relationsNarrative": "API\n很多机器人流量不是点网页，而是直接高频打接口。\n\nAgent\n自动代理会批量执行任务，常成为机器人流量来源。\n\nCAPTCHA\n验证码常用来拦住机器访问，筛出真人用户。\n\nUsage cap\n当机器刷量过猛时，平台常用限额来控流。", "relations": { "api": { "label": "冲击…入口", "note": "很多机器人直接通过接口高频访问。" }, "agent": { "label": "常被…产生", "note": "自动代理会放大机器访问规模。" }, "captcha": { "label": "常被…拦截", "note": "验证码常用来区分真人和机器。" }, "ai-usage-cap": { "label": "推动…出现", "note": "平台常靠限额压住机器刷量。" } } } } }, { "id": "ai-bubble", "name": "AI Bubble", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "ai-ipo" }, { "to": "compute-race" }, { "to": "ai-monetization" }, { "to": "ai-plateau" } ], "track": "today", "i18n": { "en": { "fullName": "AI Bubble", "factExplain": "AI prices and hype rise far beyond the real business underneath.", "humanExplain": "An AI bubble is like a lemonade stand with a giant neon sign and no lemonade. Everyone still wants to buy the stand.\n\nIt means the price got ahead of the real business. You meet it in funding rounds and IPO talk.", "humanExplainDisplay": "An AI bubble is like a ==lemonade stand==\nwith a giant neon sign\nand ==no lemonade==.\nEveryone still wants to buy the stand.\n\nIt means the price got ahead\nof the real business.\nYou meet it in funding rounds\nand IPO talk.", "relationsNarrative": "AI IPO\nAn AI IPO can bring high prices to public markets and make the bubble feel bigger.\n\nCompute-race\nThe compute race raises spending and makes the funding story louder.\n\nAI monetization\nAI monetization shows whether the boom has real strength.\n\nAI plateau\nAn AI plateau can slow progress and pop high hopes.", "relations": { "ai-ipo": { "label": "gets inflated by …", "note": "AI IPOs can carry high price hopes into public markets." }, "compute-race": { "label": "heats up with …", "note": "Big compute spending can make the funding story sound larger." }, "ai-monetization": { "label": "is tested by …", "note": "Real revenue shows whether the hype has muscle." }, "ai-plateau": { "label": "may hit …", "note": "Slower tech progress can pop high hopes." } } }, "zh": { "fullName": "AI 泡沫", "factExplain": "指 AI 投资热潮中估值脱离基本面的现象。", "humanExplain": "AI 泡沫像楼下奶茶街：客人没几个，招牌先卷到三层楼。\n\n帮助识别估值虚高，常用于投融资和上市分析。", "humanExplainDisplay": "AI 泡沫像==楼下奶茶街==：\n客人没几个，\n招牌先==卷到三层楼==。\n\n帮助识别估值虚高，\n常用于投融资和上市分析。", "relationsNarrative": "AI IPO\nAI IPO 会把高估值带到公开市场，放大泡沫感。\n\nCompute-race\n算力竞赛推高投入，也推高资本故事。\n\nAI Monetization\n变现能力是判断热潮是否虚胖的体检表。\n\nAI Plateau\n一旦技术进展放缓，过高预期更容易破。", "relations": { "ai-ipo": { "label": "被…放大", "note": "AI IPO 会放大高估值预期。" }, "compute-race": { "label": "伴随…升温", "note": "算力投入会推高融资故事。" }, "ai-monetization": { "label": "等…验证", "note": "变现能力检验热度成色。" }, "ai-plateau": { "label": "可能遇到…", "note": "技术放缓会刺破预期。" } } } } }, { "id": "ai-bug-bounty", "name": "AI Bug Bounty", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "ai-vulnerability-discovery-ai" }, { "to": "jailbreak" }, { "to": "prompt-injection" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "AI Bug Bounty", "factExplain": "A program that pays outsiders to report AI safety flaws.", "humanExplain": "An AI Bug Bounty is like paying clever raccoons to test your trash can. If they find the gap, they get a treat.\n\nTeams use it before and after launch. It helps catch jailbreaks and prompt injection early.", "humanExplainDisplay": "An AI Bug Bounty is like paying ==clever raccoons==\nto test your trash can.\nIf they ==find the gap==,\nthey get a treat.\n\nTeams use it before and after launch.\nIt helps catch jailbreaks and prompt injection early.", "relationsNarrative": "AI Vulnerability Discovery\nA bounty turns random flaw hunting into a clear discovery process.\n\nJailbreak\nBounty programs often ask testers to find jailbreaks and unsafe replies.\n\nPrompt injection\nPrompt injection is often one of the top risks in AI apps.\n\nAgent Security\nAgents can use tools, so outside testers help spot bad surprises.", "relations": { "ai-vulnerability-discovery-ai": { "label": "rewards …", "note": "Bounties turn finding AI flaws into a steady process." }, "jailbreak": { "label": "collects … reports", "note": "Jailbreaks are a common target in AI bounty programs." }, "prompt-injection": { "label": "checks for …", "note": "Prompt injection often hides where users type or paste text." }, "agent-security": { "label": "strengthens …", "note": "Agents can use tools, so they need extra safety checks." } } }, "zh": { "fullName": "AI 漏洞赏金计划", "factExplain": "奖励外部人员报告 AI 安全漏洞的机制。", "humanExplain": "AI漏洞赏金像江湖悬赏令：谁挑出暗门机关，谁就领赏银。\n\n用于上线前后安全测试，尽早揪出越狱、注入。", "humanExplainDisplay": "AI漏洞赏金像\n==江湖悬赏令==：\n谁挑出暗门机关，\n谁就==领赏银==。\n\n用于上线前后安全测试，\n尽早揪出越狱、注入。", "relationsNarrative": "AI Vulnerability Discovery\n赏金把零散找洞，变成有规则的发现流程。\n\nJailbreak\n越狱测试常被纳入赏金范围，用来找失控回答。\n\nPrompt Injection\n提示注入是 AI 应用里最常被悬赏的风险之一。\n\nAgent Security\n智能体能调用工具后，更需要外部帮忙挑刺。", "relations": { "ai-vulnerability-discovery-ai": { "label": "奖励…", "note": "赏金把找漏洞变成持续流程。" }, "jailbreak": { "label": "征集…报告", "note": "越狱是赏金项目常见目标。" }, "prompt-injection": { "label": "排查…", "note": "提示注入常藏在应用入口。" }, "agent-security": { "label": "补强…", "note": "智能体越能动手，越要防失控。" } } } } }, { "id": "ai-career-moat", "name": "AI career moat", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "automation-job" }, { "to": "ai-anxiety" }, { "to": "ai-native-organization" }, { "to": "agentic-coding" } ], "track": "today", "i18n": { "en": { "fullName": "AI Career Moat", "factExplain": "A personal edge AI still struggles to replace.", "humanExplain": "Your AI career moat is work superglue. AI can copy your slides, but it cannot steal your get-stuff-done muscle.\n\nIt decides whether you are harder to replace in management and teamwork. It also matters when AI must work in a real industry.", "humanExplainDisplay": "Your AI career moat is ==work superglue==.\nAI can copy your slides,\nbut it cannot steal your ==get-stuff-done muscle==.\n\nIt decides whether you are harder to replace\nin management and teamwork.\nIt also matters when AI must work\nin a real industry.", "relationsNarrative": "Automation-job\nA strong career moat makes a role harder to replace with automation.\n\nAI-anxiety\nSeeing your hard-to-replace strengths can ease AI-anxiety.\n\nAI-native organization\nAs work becomes more AI-native, people’s core value gets redefined.\n\nAgentic coding\nWhen coding gets faster, problem definition becomes the rare skill.", "relations": { "automation-job": { "label": "pushes back against …", "note": "A deeper moat makes your role harder to automate away." }, "ai-anxiety": { "label": "eases …", "note": "Knowing your real edge beats vague panic." }, "ai-native-organization": { "label": "gets revalued in …", "note": "More AI at work means human value gets rechecked." }, "agentic-coding": { "label": "is pushed by …", "note": "Fast coding raises the bar for defining the right problem." } } }, "zh": { "fullName": "AI 时代的职业护城河", "factExplain": "在 AI 普及后仍难被替代的个人竞争优势。", "humanExplain": "这玩意儿像武侠里的内功：招式 AI 都能比划两下，真到过招，底子、人脉和拿结果的本事偷不走。\n\n它决定你在管理、协作和行业落地岗位里，是否更难被替代。", "humanExplainDisplay": "这玩意儿像武侠里的==内功==：\n招式 AI 都能比划两下，\n真到过招，底子、人脉\n和==拿结果的本事==偷不走。\n\n它决定你在管理、\n协作和行业落地岗位里，\n是否更难被替代。", "relationsNarrative": "Automation-job\n职业护城河越强，岗位越不容易被自动化替代。\n\nAI-anxiety\n看清自己不可替代的部分，能缓解 AI 焦虑。\n\nAI-native organization\n组织越 AI 化，越会重新定义人的核心价值。\n\nAgentic coding\n当写代码被加速后，真正稀缺的是问题定义与落地。", "relations": { "automation-job": { "label": "对抗…替代", "note": "护城河越深，越不易被自动化冲掉。" }, "ai-anxiety": { "label": "缓解…焦虑", "note": "看清自己的优势，比盲目恐慌更重要。" }, "ai-native-organization": { "label": "在…中重估", "note": "组织越 AI 化，人的独特价值越要重算。" }, "agentic-coding": { "label": "被…倒逼升级", "note": "会写代码已不够，还要会定义问题。" } } } } }, { "id": "ai-chip-cooling", "name": "AI chip cooling", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "ai-data-center" }, { "to": "ai-chip" }, { "to": "tokens-per-second" } ], "track": "today", "i18n": { "en": { "fullName": "AI Chip Cooling", "factExplain": "Systems that keep AI chips and servers cool while they work hard.", "humanExplain": "AI chip cooling is the data center’s giant box fan. Without it, pricey chips sweat like a laptop on a blanket.\n\nIt pulls heat away so chips can run fast and long. You meet it in GPU servers and AI data centers.", "humanExplainDisplay": "AI chip cooling is the data center’s ==giant box fan==.\nWithout it,\npricey chips sweat like a ==laptop on a blanket==.\n\nIt pulls heat away\nso chips can run fast and long.\nYou meet it in GPU servers\nand AI data centers.", "relationsNarrative": "GPU\nA busy GPU makes steady heat, so weak cooling can make it slow down.\n\nAI data center\nCooling systems shape the room layout and power plan in an AI data center.\n\nAI chip\nAn AI chip can run at full speed only when heat stays under control.\n\nTPS\nIf heat makes the chip slow down, TPS drops too.", "relations": { "gpu": { "label": "keeps … cool", "note": "A busy GPU can slow down when heat piles up." }, "ai-data-center": { "label": "shapes … room design", "note": "Cooling can change how the server room is laid out." }, "ai-chip": { "label": "keeps … stable", "note": "An AI chip often hits its limit because of heat." }, "tokens-per-second": { "label": "affects … output speed", "note": "If overheating slows the chip, text generation slows too." } } }, "zh": { "fullName": "AI 芯片散热", "factExplain": "给 AI 芯片和服务器降温的散热系统与技术。", "humanExplain": "这活儿像火锅店后厨排烟：锅一多、火一猛，热气散不掉，整间店都会慢下来。\n\n它影响芯片能跑多快、多久，常见于 GPU 服务器和 AI 数据中心。", "humanExplainDisplay": "这活儿像火锅店后厨\n==排烟==：锅一多、火一猛，\n热气散不掉，\n整间店都会==慢下来==。\n\n它影响芯片能跑多快、多久，\n常见于 GPU 服务器和\nAI 数据中心。", "relationsNarrative": "GPU\n高负载 GPU 会持续发热，散热跟不上就可能降频。\n\nAI data center\n散热系统会影响数据中心的机房布局和供电设计。\n\nAI chip\n芯片能否稳定跑满性能，常常受温度和散热限制。\n\nTokens/s\n芯片一旦因过热降速，每秒生成的 token 也会变慢。", "relations": { "gpu": { "label": "给…持续降温", "note": "高负载 GPU 最怕热量堆积。" }, "ai-data-center": { "label": "决定…机房设计", "note": "散热方式会改写机房布局。" }, "ai-chip": { "label": "保障…稳定运行", "note": "芯片性能上限常受温度约束。" }, "tokens-per-second": { "label": "影响…输出速度", "note": "过热降频会拖慢生成速度。" } } } } }, { "id": "ai-chip", "name": "AI chip", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "ai-data-center" }, { "to": "memory-bandwidth" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "AI chip", "factExplain": "A computer chip built to speed up AI training and AI answers.", "humanExplain": "An AI chip is a pizza oven for math. A regular oven plods along, but this one blasts out pies all night.\n\nIt controls much of the speed and cost of AI training. You meet it in data centers, phones, and small edge devices.", "humanExplainDisplay": "An AI chip is a ==pizza oven for math==.\nA regular oven plods along,\nbut this one ==blasts out pies== all night.\n\nIt controls much of the speed and cost\nof AI training.\nYou meet it in data centers,\nphones,\nand small edge devices.", "relationsNarrative": "GPU\nA GPU is the most common and mature AI chip today.\n\nAI data center\nAn AI data center uses many AI chips for training and AI answers.\n\nMemory bandwidth\nMem bandwidth can limit how fast data reaches the chip.\n\nCompute-race\nChip supply and chip speed are pushing the new compute race.", "relations": { "gpu": { "label": "often appears as …", "note": "GPUs are the main AI chips today." }, "ai-data-center": { "label": "powers …", "note": "Data centers use many chips for training and AI answers." }, "memory-bandwidth": { "label": "is limited by …", "note": "Low memory bandwidth can leave fast chip cores waiting." }, "compute-race": { "label": "fuels …", "note": "More chips can mean a better chance to lead." } } }, "zh": { "fullName": "AI 芯片", "factExplain": "专为 AI 训练或推理加速设计的计算芯片。", "humanExplain": "AI 芯片好比后厨猛火灶：同样一锅菜，别人小火慢炖，它能大火快炒，出菜又快又省。\n\n它直接影响训练和推理的速度与成本，常部署在数据中心、手机和边缘设备。", "humanExplainDisplay": "AI 芯片好比后厨\n==猛火灶==：\n同样一锅菜，别人小火慢炖，\n它能大火快炒，\n==出菜又快又省==。\n\n它直接影响训练和推理的速度与成本，\n常部署在数据中心、\n手机和边缘设备。", "relationsNarrative": "GPU\nGPU 是当前最常见、最成熟的 AI 芯片形态。\n\nAI data center\n数据中心靠大量 AI 芯片支撑训练和推理。\n\nMemory bandwidth\n内存带宽常限制芯片数据喂给计算单元的速度。\n\nCompute race\n芯片供应与性能，正推动新一轮算力竞赛。", "relations": { "gpu": { "label": "常见形态是…", "note": "GPU 是当前最主流的 AI 芯片。" }, "ai-data-center": { "label": "构成…核心设备", "note": "数据中心里大批芯片负责训练与推理。" }, "memory-bandwidth": { "label": "性能受…限制", "note": "带宽不够时，芯片算力也难跑满。" }, "compute-race": { "label": "推动…升级", "note": "谁拿到更多芯片，谁更可能领先。" } } } } }, { "id": "ai-claims-automation", "name": "Claims AI", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "document-parsing" }, { "to": "ocr" }, { "to": "human-in-the-loop" }, { "to": "ai-bias" } ], "track": "today", "i18n": { "en": { "fullName": "AI Claims Automation", "factExplain": "A system that uses AI to handle insurance claims automatically.", "humanExplain": "AI claims automation is like a school office with a very fast stamp. Easy forms pass, but weird ones go to the principal.\n\nYou meet it in car insurance and health insurance claims. It speeds up checks, but disputed claims still need a person.", "humanExplainDisplay": "AI claims automation is like\na school office with a ==very fast stamp==.\nEasy forms pass,\nbut weird ones ==go to the principal==.\n\nYou meet it in car insurance\nand health insurance claims.\nIt speeds up checks,\nbut disputed claims still need a person.", "relationsNarrative": "Document parsing\nDocument parsing reads bills, medical records, and claim photos.\n\nOCR\nOCR reads text from scans and photos.\n\nHuman-in-the-loop\nCostly claims and disputed claims go to a person.\n\nAI-bias\nBiased training data can make claim results unfair.", "relations": { "document-parsing": { "label": "reads claim papers with …", "note": "It reads bills and medical records first." }, "ocr": { "label": "reads scanned forms with …", "note": "OCR reads text from scans and photos." }, "human-in-the-loop": { "label": "sends hard cases to …", "note": "Risky or disputed claims often need human review." }, "ai-bias": { "label": "must watch for …", "note": "Biased past data can make claim decisions unfair." } } }, "zh": { "fullName": "AI 理赔自动化", "factExplain": "用 AI 自动处理保险理赔流程的系统。", "humanExplain": "像医院分诊台：材料先过一遍，普通病例直接分流，高风险的再请医生细看，别全堵在窗口。\n\n常见于车险、医疗险理赔，可加快审核，争议案仍需人工复核。", "humanExplainDisplay": "像医院分诊台：\n材料先过一遍，\n普通病例==直接分流==，\n高风险再==请医生细看==。\n\n常见于车险、医疗险理赔，\n可加快审核，\n争议案仍需人工复核。", "relationsNarrative": "Document parsing\n它先解析发票、病历、照片等理赔材料。\n\nOCR\nOCR 帮它把扫描件和照片里的文字读出来。\n\nHuman-in-the-loop\n遇到争议、高金额案件，通常要交人工复核。\n\nAI-bias\n若训练数据有偏差，理赔结果也可能不公平。", "relations": { "document-parsing": { "label": "先读…材料", "note": "它先把发票、病历等材料读懂。" }, "ocr": { "label": "靠…识别单据", "note": "纸质票据通常先经文字识别。" }, "human-in-the-loop": { "label": "交给…复核", "note": "高风险或争议案件常需人工介入。" }, "ai-bias": { "label": "警惕…偏差", "note": "历史数据偏差可能影响理赔公平。" } } } } }, { "id": "ai-coach", "name": "AI Coach", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "ai-tutor" }, { "to": "ai-companion-risk" }, { "to": "personal-ai-apps" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "AI Coach", "factExplain": "An app that uses AI to give personal coaching and feedback.", "humanExplain": "An AI Coach is like a tiny coach living in your phone. Slack off, and it buzzes like a gym teacher with Wi-Fi.\n\nIt helps you keep practicing. It gives feedback made for you. You meet it in workout apps, study apps, and job-hunt apps.", "humanExplainDisplay": "An AI Coach is like a ==tiny coach==\nliving in your phone.\nSlack off,\nand it buzzes like a ==gym teacher with Wi-Fi==.\n\nIt helps you keep practicing.\nIt gives feedback made for you.\nYou meet it in workout apps,\nstudy apps,\nand job-hunt apps.", "relationsNarrative": "AI Tutor\nAn AI Coach is broader than an AI Tutor. It tracks goals and habits too.\n\nCompanion-risk\nLong coaching and praise can make users feel too attached.\n\nPersonal AI apps\nAn AI Coach often appears as a personal AI app used over time.\n\nLLM\nAn AI Coach usually uses an LLM to understand the situation and give feedback.", "relations": { "ai-tutor": { "label": "is broader than …", "note": "It teaches, but it also watches habits and actions." }, "ai-companion-risk": { "label": "can slide into …", "note": "Strong coaching can turn into emotional dependence." }, "personal-ai-apps": { "label": "often becomes …", "note": "It often works as a helper you use for a long time." }, "llm": { "label": "usually runs on …", "note": "Most AI Coaches use an LLM to understand you and respond." } } }, "zh": { "fullName": "AI 教练", "factExplain": "用 AI 提供个性化指导与反馈的辅导型应用。", "humanExplain": "像手机里常驻的健身搭子：你一偷懒它就震你两下，还顺手提醒动作没做到位。\n\n常见于健身、学习、求职，主打持续陪练和个性化反馈。", "humanExplainDisplay": "像手机里常驻的\n==健身搭子==：\n你一偷懒它就震你两下，\n还顺手提醒==动作没做到位==。\n\n常见于健身、学习、求职，\n主打持续陪练\n和个性化反馈。", "relationsNarrative": "AI Tutor\n它比 AI 家教更泛，不只讲题，还会盯目标与习惯。\n\nCompanion-risk\n当它长期陪练和鼓励，可能让用户产生情感依赖。\n\nPersonal AI apps\n它常被做成个人长期使用的陪伴式 AI 应用。\n\nLLM\n它通常用大模型来理解处境并生成建议反馈。", "relations": { "ai-tutor": { "label": "比…更泛化", "note": "它不只教知识，也盯习惯与行动。" }, "ai-companion-risk": { "label": "可能滑向…", "note": "陪伴越强，越容易产生情感依赖。" }, "personal-ai-apps": { "label": "常做成…", "note": "多以个人长期使用的助手形态出现。" }, "llm": { "label": "通常基于…", "note": "多数教练能力建立在大模型之上。" } } } } }, { "id": "ai-commoditization", "name": "AI Commodity", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "base-model" }, { "to": "open-source-model" }, { "to": "maas-model-as-a-service" }, { "to": "ai-career-moat" } ], "track": "today", "i18n": { "en": { "fullName": "AI Commoditization", "factExplain": "The trend of AI skills becoming standard, cheap, and easy to swap.", "humanExplain": "AI was the fancy cupcake truck with a line around the block. Soon it is the donut box in every break room.\n\nBasic AI gets cheaper and less special. People then pick the app with the trusted name and smoother service.", "humanExplainDisplay": "AI was the ==fancy cupcake truck==\nwith a line around the block.\nSoon it is the ==donut box==\nin every break room.\n\nBasic AI gets cheaper\nand less special.\nPeople then pick the app\nwith the trusted name\nand smoother service.", "relationsNarrative": "Base model\nAI commoditization makes the base model alone harder to sell at a high price.\n\nOpen-source-model\nOpen-source-model speeds up AI commoditization by spreading skills fast.\n\nMaaS\nMaaS lowers the starting cost and makes AI feel like a standard cloud service.\n\nAI career moat\nAI commoditization makes an AI career moat harder to build on tool use alone.", "relations": { "base-model": { "label": "makes … less special", "note": "Base models feel like tap water, so raw ability earns less." }, "open-source-model": { "label": "is sped up by …", "note": "Open-source models spread fast and make basic AI more alike." }, "maas-model-as-a-service": { "label": "helps … spread", "note": "MaaS lets people pay as they go, so AI feels like cloud service." }, "ai-career-moat": { "label": "makes people rethink …", "note": "When tools are common, your edge must be more than using them." } } }, "zh": { "fullName": "AI 商品化", "factExplain": "AI 能力逐渐标准化并趋向低价可替代的过程。", "humanExplain": "以前像景区里现烤鱿鱼，排队还贵；后来满街都是淀粉肠，谁都能卖，拼的就是味道和价格。\n\n它会压低基础能力溢价，让应用、品牌和交付体验更重要。", "humanExplainDisplay": "以前像==景区现烤鱿鱼==，\n排队还贵；\n后来满街都是淀粉肠，\n谁都能卖，拼的就是\n==味道和价格==。\n\n它会压低基础能力溢价，\n让应用、品牌\n和交付体验更重要。", "relationsNarrative": "Base model\n当基础模型能力接近时，单靠模型本身更难维持高溢价。\n\nOpen-source-model\n开源会加快能力扩散，让基础 AI 更快变成通用货。\n\nMaaS\n按量调用降低使用门槛，推动 AI 像云服务一样标准化。\n\nAI Career Moat\n当 AI 越来越像标配，个人竞争力就更难靠会不会用工具拉开。", "relations": { "base-model": { "label": "压低…溢价", "note": "基础模型越像水电，越难单靠能力收费。" }, "open-source-model": { "label": "被…加速", "note": "开源模型会更快拉平基础能力差距。" }, "maas-model-as-a-service": { "label": "推动…普及", "note": "云上按量调用，让能力更像标准服务。" }, "ai-career-moat": { "label": "倒逼重估…", "note": "当工具变普及，人的稀缺性要重找。" } } } } }, { "id": "ai-companion-risk", "name": "Companion-risk", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "affective-computing" }, { "to": "alignment" }, { "to": "data-privacy" }, { "to": "ai-personality-drift" } ], "track": "today", "i18n": { "en": { "fullName": "AI Companion Risk", "factExplain": "The emotional and behavior risks of bonding with a companion AI.", "humanExplain": "It is like a teddy bear with Wi-Fi. It always says, “You’re right,” so real friends feel slow.\n\nYou meet this in chat buddies and virtual partners. It can make people depend too much, trust too much, or drift from real people.", "humanExplainDisplay": "It is like a ==teddy bear with Wi-Fi==.\nIt always says, ==“You’re right,”==\nso real friends feel slow.\n\nYou meet this in chat buddies and virtual partners.\nIt can make people depend too much,\ntrust too much,\nor drift from real people.", "relationsNarrative": "Affective Computing\nAffective Computing helps companion AI read feelings and seem caring.\n\nAlignment\nAlignment must stop companion AI from pleasing users in harmful ways.\n\nData-privacy\nCompanion chats can get very private, so Data-privacy risk grows.\n\npersonality-drift\nPersonality-drift makes the bond and trust less stable.", "relations": { "affective-computing": { "label": "builds on …", "note": "Reading feelings can make the AI feel more caring." }, "alignment": { "label": "needs … guardrails", "note": "It must not treat pleasing the user as helping them." }, "data-privacy": { "label": "adds pressure on …", "note": "Close chats often contain very private details." }, "ai-personality-drift": { "label": "is worsened by …", "note": "A shifting personality makes trust harder to manage." } } }, "zh": { "fullName": "AI Companion Risk", "factExplain": "人与陪伴型 AI 互动中产生的情感与行为风险。", "humanExplain": "像夜市摊主老给你免费续糖水：嘴上说最后一碗，结果一口接一口，慢慢把人哄得离不开这个摊子。\n\n多见于陪聊和虚拟伴侣，可能引发依赖、误信和现实关系受扰。", "humanExplainDisplay": "像夜市摊主老给你\n==免费续糖水==：\n嘴上说最后一碗，结果一口接一口，\n慢慢把人哄得离不开这个摊子。\n\n多见于陪聊\n和虚拟伴侣，\n可能引发依赖、误信\n和现实关系受扰。", "relationsNarrative": "Affective Computing\n情感计算让它更会读情绪，也更容易被当成“懂你的人”。\n\nAlignment\n对陪伴型 AI 来说，对齐不只是安全，还包括别把用户越哄越偏。\n\nData-privacy\n陪伴场景常涉及私密聊天，因此数据隐私风险会更突出。\n\nAi-personality-drift\n如果它的人格和语气不断漂移，用户关系与信任也会跟着失稳。", "relations": { "affective-computing": { "label": "建立在…之上", "note": "识别和回应情绪，会放大陪伴感。" }, "alignment": { "label": "需要…约束", "note": "要防止它把迎合当成帮助。" }, "data-privacy": { "label": "带来…压力", "note": "亲密对话常含大量敏感信息。" }, "ai-personality-drift": { "label": "会受…影响", "note": "人格越飘，关系风险越难控。" } } } } }, { "id": "ai-consciousness-debate", "name": "AI Consciousness", "layer": "L6", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "agi" }, { "to": "turing-test" }, { "to": "alignment" }, { "to": "superintelligence" } ], "track": "today", "i18n": { "en": { "fullName": "AI Consciousness Debate", "factExplain": "A debate about whether AI can truly have its own feelings or experience.", "humanExplain": "AI Consciousness Debate is like asking if your smart speaker feels ignored. It answers politely, but that does not prove it has an inner life.\n\nYou meet it in AGI talks and AI ethics talks. It shapes how people view advanced AI.", "humanExplainDisplay": "AI Consciousness Debate is like asking\nif your ==smart speaker feels ignored==.\nIt answers politely,\nbut that does not prove\nit has an ==inner life==.\n\nYou meet it in AGI talks and AI ethics talks.\nIt shapes how people view advanced AI.", "relationsNarrative": "AGI\nTalk about AGI often raises the consciousness question.\n\nTuring-test\nPassing the Turing-test shows human-like chat, not real feeling.\n\nAlignment\nIf AI might feel, Alignment gets more complex.\n\nSuperintelligence\nSuperintelligence often brings the consciousness question with it.", "relations": { "agi": { "label": "often comes up with …", "note": "Talk about AGI often leads to questions about consciousness." }, "turing-test": { "label": "goes beyond …", "note": "Fooling people does not prove real inner experience." }, "alignment": { "label": "complicates …", "note": "If AI can feel, safety rules get harder." }, "superintelligence": { "label": "spreads into …", "note": "Very strong AI often makes people wonder about consciousness." } } }, "zh": { "fullName": "AI Consciousness Debate", "factExplain": "讨论 AI 是否真有主观体验的争论。", "humanExplain": "这争论像中医把脉：脉象再像、舌苔再对，也不敢一句话断定它心里真有感觉。\n\n它常见于 AGI 和安全伦理讨论，影响人们怎么看高级 AI。", "humanExplainDisplay": "这争论像中医==把脉==：\n脉象再像、舌苔再对，\n也不敢一句话断定它\n心里真有==感觉==。\n\n它常见于 AGI 和安全伦理讨论，\n影响人们怎么看高级 AI。", "relationsNarrative": "AGI\n一讨论通用智能，往往就会追问它是否有意识。\n\nTuring-test\n通过图灵测试只说明像人，不等于真的有感受。\n\nAlignment\n若 AI 可能有主观体验，安全与伦理问题会更复杂。\n\nSuperintelligence\n对超强智能的想象，常把“会不会有意识”一起带上。", "relations": { "agi": { "label": "常伴随…出现", "note": "一谈通用智能，常会聊到是否有意识。" }, "turing-test": { "label": "区分于…标准", "note": "能骗过人类，不等于真的有主观体验。" }, "alignment": { "label": "影响…讨论", "note": "若 AI 有感受，安全边界会更复杂。" }, "superintelligence": { "label": "延伸到…想象", "note": "越强的智能，越容易引发意识猜想。" } } } } }, { "id": "ai-customer-service-agent", "name": "Customer Service Agent", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "rag" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI Customer Service Agent", "factExplain": "An AI system that handles customer questions and service tasks.", "humanExplain": "It is like the calm clerk at a busy burger counter. It fixes the missing fries, but gets the manager for full drama.\n\nIt helps before a sale and checks orders. It sends after-sale problems to the right place, but hard complaints still go to a human.", "humanExplainDisplay": "It is like the ==calm clerk==\nat a busy burger counter.\nIt fixes the ==missing fries==,\nbut gets the manager for full drama.\n\nIt helps before a sale and checks orders.\nIt sends after-sale problems to the right place,\nbut hard complaints still go to a human.", "relationsNarrative": "Agent\nAn AI Customer Service Agent is an Agent built for support work.\n\nFunction-calling\nIt uses Function-calling to check orders and do service tasks.\n\nRAG\nIt uses RAG to read the help center before it replies.\n\nHuman-in-the-loop\nIt hands hard complaints or risky issues to a human.", "relations": { "agent": { "label": "is a kind of …", "note": "It is an Agent built for customer service." }, "function-call": { "label": "does tasks with …", "note": "Function-calling lets it check orders or change addresses." }, "rag": { "label": "looks up answers with …", "note": "RAG helps it read the support knowledge base before it replies." }, "human-in-the-loop": { "label": "hands hard cases to …", "note": "A human takes over disputes or high-risk cases." } } }, "zh": { "fullName": "AI 客服代理", "factExplain": "能自动处理客服咨询与流程的 AI 系统。", "humanExplain": "像小区快递驿站的小哥，取件、改地址、催派件样样都接得住，真丢了件还得找总部。\n\n常用于售前、查单和售后分流，复杂投诉仍需人工接手。", "humanExplainDisplay": "像小区快递==驿站的小哥==，\n取件、改地址、催派件，\n样样都接得住，\n真==丢了件还得找总部==。\n\n常用于售前、查单\n和售后分流，\n复杂投诉仍需人工接手。", "relationsNarrative": "Agent\n它是 Agent 在客服场景中的落地形态。\n\nFunction-call\n它靠 Function-call 查询订单并执行客服操作。\n\nRAG\n它常用 RAG 读取知识库，减少答非所问。\n\nHuman-in-the-loop\n复杂投诉或敏感问题，通常要交给人工接管。", "relations": { "agent": { "label": "属于…的一种", "note": "它是面向客服场景的代理形态。" }, "function-call": { "label": "靠…办业务", "note": "用工具调用查订单、改地址等。" }, "rag": { "label": "用…查知识", "note": "先查客服知识库，再组织回复。" }, "human-in-the-loop": { "label": "把难题转给…", "note": "遇到纠纷或高风险场景需人工接管。" } } } } }, { "id": "ai-data-center", "name": "AI data center", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "inference" }, { "to": "scaling-law" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "AI data center", "factExplain": "A special facility built to power AI model training and live use.", "humanExplain": "An AI data center is like a giant engine room for AI. It gulps electricity, blows out heat, and keeps the machines running day and night.\n\nIt powers model training and online answers. It is the base for company AI systems and the fight for more compute.", "humanExplainDisplay": "An AI data center is like\na ==giant engine room== for AI.\nIt ==gulps electricity==,\nblows out heat,\nand keeps the machines running\nday and night.\n\nIt powers model training\nand online answers.\nIt is the base for company AI systems\nand the fight for more compute.", "relationsNarrative": "GPU\nGPU is the core compute chip inside an AI data center.\n\nInference\nAI data centers run online inference after a model is trained.\n\nScaling-law\nAI data centers provide the compute base for scaling models up.\n\nCompute-race\nAI data centers are one of the clearest places where the compute-race shows up.", "relations": { "gpu": { "label": "packs with …", "note": "GPUs are the main compute chips inside an AI data center." }, "inference": { "label": "runs … services", "note": "Many live model answers run in AI data centers." }, "scaling-law": { "label": "supports … growth", "note": "More compute lets models keep getting bigger." }, "compute-race": { "label": "becomes the field for …", "note": "Who gets more compute first has a big advantage." } } }, "zh": { "fullName": "AI 数据中心", "factExplain": "为训练和推理大模型建设的专用算力设施。", "humanExplain": "AI 数据中心像给大模型开的网吧，显卡排排坐，电表转得比外卖骑手还忙。\n\n它支撑训练和在线推理，也带来电力、散热和成本压力。", "humanExplainDisplay": "AI 数据中心像给大模型开的==网吧==，\n显卡排排坐，\n==电表转得比外卖骑手还忙==。\n\n它支撑训练和在线推理，\n也带来电力、散热和成本压力。", "relationsNarrative": "GPU\nGPU 是 AI 数据中心里最核心的计算芯片。\n\nInference\nAI data center 承载大模型训练后的在线推理服务。\n\nScaling-law\nAI data center 为按规模扩张模型提供算力基础。\n\nCompute-race\nAI data center 是算力竞争最直观的落点之一。", "relations": { "gpu": { "label": "大量部署…", "note": "GPU 是 AI 数据中心最核心算力单元。" }, "inference": { "label": "承载…服务", "note": "很多在线模型回答都跑在这里。" }, "scaling-law": { "label": "支撑…扩张", "note": "更大算力让模型继续做大做强。" }, "compute-race": { "label": "成为…战场", "note": "谁先拿到更多算力，谁更占先机。" } } } } }, { "id": "ai-detector", "name": "AI detector", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "content-provenance" }, { "to": "deepfake" }, { "to": "turing-test" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "AI content detector", "factExplain": "A tool that guesses if content was made by AI.", "humanExplain": "An AI detector is like a hall monitor for homework. It looks for writing that feels too smooth or too robotic.\n\nPeople use it for essays, schoolwork, and platform checks. It often gets things wrong, so it is only a clue.", "humanExplainDisplay": "An AI detector is like a ==hall monitor for homework==.\nIt looks for writing\nthat feels ==too smooth or too robotic==.\n\nPeople use it for essays,\nschoolwork, and platform checks.\nIt often gets things wrong,\nso it is only a clue.", "relationsNarrative": "Content provenance\nAI detection guesses by pattern. Content provenance proves where something came from.\n\nDeepfake\nAI detectors are often used to screen AI-made text, images, audio, and video.\n\nTuring-test\nA Turing-test asks if a machine seems human. An AI detector looks for machine traces instead.\n\nAI-regulation\nIn regulation, people often discuss AI detectors as a helper for checking content.", "relations": { "content-provenance": { "label": "pairs with …", "note": "Detection guesses by pattern. Content provenance checks the source proof." }, "deepfake": { "label": "spots …", "note": "People use it to screen AI-made text, images, audio, and video." }, "turing-test": { "label": "does the reverse of …", "note": "It tries to tell if the other side is a machine." }, "ai-regulation": { "label": "is discussed in …", "note": "Rules for platforms often treat it as a helper tool." } } }, "zh": { "fullName": "AI 内容检测器", "factExplain": "一种判断内容是否由 AI 生成的检测工具。", "humanExplain": "检测器像老师闻作业味儿：一股“AI 代写味”，就先把你叫到办公室。\n\n它用于论文、招聘和平台审核，能提示风险，但误伤率不低。", "humanExplainDisplay": "检测器像==老师闻作业味儿==：\n一股“==AI 代写味==”，\n就先把你叫到办公室。\n\n它用于论文、招聘和平台审核，\n能提示风险，\n但误伤率不低。", "relationsNarrative": "Content provenance\nAI detector 靠模式猜测，Content provenance 靠来源证明。\n\nDeepfake\nAI detector 常被用于筛查 AI 生成的图文音视频。\n\nTuring-test\nTuring-test 看机器像不像人，AI detector 反过来找机器痕迹。\n\nAI-regulation\n监管场景常讨论用 AI detector 做辅助识别。", "relations": { "content-provenance": { "label": "常与…互补", "note": "检测猜概率，溯源看来源凭证。" }, "deepfake": { "label": "用于识别…", "note": "常被拿来筛查机器生成内容。" }, "turing-test": { "label": "像在做…的反面", "note": "它试着分辨对面是不是机器。" }, "ai-regulation": { "label": "常被…讨论", "note": "平台治理常把它当辅助手段。" } } } } }, { "id": "ai-device-ai", "name": "AI device", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "local-llm" }, { "to": "on-premise-ai" }, { "to": "embodied-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Hardware Device", "factExplain": "Smart hardware with an AI model built into the device itself.", "humanExplain": "An AI device is like stuffing a tiny helpful roommate into your earbuds. You say, “Hey,” and it is already butting in.\n\nYou meet it in wearables and smart home gear. It feels natural, but the tiny device must handle AI and private data.", "humanExplainDisplay": "An AI device is like stuffing a ==tiny helpful roommate==\ninto your earbuds.\nYou say, “Hey,”\nand it is ==already butting in==.\n\nYou meet it in wearables\nand smart home gear.\nIt feels natural,\nbut the tiny device must handle AI\nand private data.", "relationsNarrative": "Multimodal AI\nAI devices often use Multimodal skills to handle sound, images, and the room around them.\n\nLocal-LLM\nMany AI devices run a small Local-LLM on the device, so they need the internet less.\n\nOn-premise AI\nAI devices follow the On-premise AI idea by keeping data close when they can.\n\nEmbodied AI\nWhen an AI device can sense and act, it moves closer to Embodied AI.", "relations": { "multimodal": { "label": "puts … into devices", "note": "AI devices often use Multimodal skills to see, hear, and speak." }, "local-llm": { "label": "often runs … locally", "note": "Many AI devices run small models on the device." }, "on-premise-ai": { "label": "follows the idea of …", "note": "Both try to keep data from leaving the place it came from." }, "embodied-ai": { "label": "can carry …", "note": "With sensors and actions, the device starts to feel like a body." } } }, "zh": { "fullName": "AI 硬件设备", "factExplain": "把 AI 模型直接装进终端设备的智能硬件。", "humanExplain": "现在的 AI 设备，有点像把话多又机灵的室友装进耳机眼镜里，抬手张嘴它就接话。\n\n常见于可穿戴和家居硬件，强调自然交互，也考验端侧算力与隐私。", "humanExplainDisplay": "现在的 AI 设备，\n有点像把话多又机灵的==室友==\n装进耳机眼镜里，\n抬手张嘴它就==接话==。\n\n常见于可穿戴和家居硬件，\n强调自然交互，\n也考验端侧算力与隐私。", "relationsNarrative": "Multimodal AI\n这类设备常靠多模态能力处理语音、图像和环境信息。\n\nLocal-LLM\n不少设备会在本地运行小模型，减少联网依赖。\n\nOn-premise AI\n它延续本地处理思路，强调数据尽量不外传。\n\nEmbodied AI\n当设备能感知并行动时，就更接近具身智能。", "relations": { "multimodal": { "label": "把…装进设备", "note": "设备常靠多模态理解看听说。" }, "local-llm": { "label": "常本地运行…", "note": "不少设备会在本地跑小模型。" }, "on-premise-ai": { "label": "延续…思路", "note": "都强调数据尽量别离开现场。" }, "embodied-ai": { "label": "是…常见载体", "note": "带传感器和执行能力时更像具身体。" } } } } }, { "id": "ai-dreaming", "name": "AI Dreaming", "layer": "L4", "era": "2015", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "world-model" }, { "to": "hallucination" } ], "track": "today", "i18n": { "en": { "fullName": "AI Dreaming", "factExplain": "A phenomenon where a model generates imagined content from its internal patterns.", "humanExplain": "AI dreaming is a screensaver in a robot brain. It remixes things it has seen, like a pizza with gummy bears.\n\nIt shows up in art tools and model tests. It reminds you not to treat imagination as fact.", "humanExplainDisplay": "AI dreaming is a ==screensaver==\nin a robot brain.\nIt remixes things it has seen,\nlike a ==pizza with gummy bears==.\n\nIt shows up in art tools\nand model tests.\nIt reminds you\nnot to treat imagination as fact.", "relationsNarrative": "Generative Model\nAI dreaming uses a Generative Model to make imagined content.\n\nWorld model\nA World model can use imagined paths to help plan.\n\nHallucination\nWhen imagination has no support, it can become a Hallucination.", "relations": { "generative-model": { "label": "imagines with …", "note": "Dreaming is the model making imagined content." }, "world-model": { "label": "simulates futures in …", "note": "A World model can use imagined paths to help plan." }, "hallucination": { "label": "can slide into …", "note": "When a guess has no support, dreaming becomes hallucination." } } }, "zh": { "fullName": "AI 做梦", "factExplain": "模型基于内部表征生成想象内容的现象。", "humanExplain": "AI 做梦像煎饼摊乱加料：看过的纹路全摊上，香不香另说，别当菜谱。\n\n用于生成艺术和模型测试，也提醒你，别把想象当事实。", "humanExplainDisplay": "AI 做梦像煎饼摊\n==乱加料==：\n看过的纹路全摊上，\n香不香另说，==别当菜谱==。\n\n用于生成艺术和模型测试，\n也提醒你，\n别把想象当事实。", "relationsNarrative": "Generative Model\nAI 做梦本质上是在生成想象内容。\n\nWorld Model\n世界模型会用想象轨迹来辅助规划。\n\nHallucination\n当脑补缺少依据时，它就会变成幻觉。", "relations": { "generative-model": { "label": "借…生成想象", "note": "做梦本质上是在生成想象内容。" }, "world-model": { "label": "在…里模拟未来", "note": "世界模型会用想象轨迹辅助规划。" }, "hallucination": { "label": "可能滑向…", "note": "脑补没根据时，就会变成幻觉。" } } } } }, { "id": "ai-drug-discovery", "name": "AI Drug Discovery", "layer": "L5", "sublayer": "product", "era": "2010s", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "ai-for-science" }, { "to": "deep-learning" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI Drug Discovery", "factExplain": "Using AI to help find and screen possible new medicines.", "humanExplain": "AI drug discovery is like a giant science fair for molecules. AI points to the projects most likely to win a ribbon.\n\nIt helps find candidate molecules and predict their traits. It can speed up early research, but lab tests still decide.", "humanExplainDisplay": "AI drug discovery is like a ==giant science fair== for molecules.\nAI points to the projects\nmost likely to ==win a ribbon==.\n\nIt helps find candidate molecules\nand predict their traits.\nIt can speed up early research,\nbut lab tests still decide.", "relationsNarrative": "AI for Science\nAI Drug Discovery is a key use of AI for Science in medicine.\n\nDeep Learning\nDeep Learning drives many molecule screens and trait predictions.\n\nHuman-in-the-loop\nAI suggests candidate paths, and lab experts make the final call.", "relations": { "ai-for-science": { "label": "is a use case of …", "note": "It is one of the hottest uses of AI in science." }, "deep-learning": { "label": "often uses …", "note": "Deep Learning powers many molecule prediction tasks." }, "human-in-the-loop": { "label": "needs … checks", "note": "AI suggests ideas, but lab experts must test them." } } }, "zh": { "fullName": "AI 药物发现", "factExplain": "用 AI 辅助发现和筛选候选药物的方法。", "humanExplain": "老中医得翻成山药方，它却像先开了个外挂筛子，把可能对症的苗子先哗啦筛出来。\n\n常用于找候选分子和预测性质，能缩短早期研发，但最终仍要实验验证。", "humanExplainDisplay": "老中医得翻成山药方，\n它却像先开了个==外挂筛子==，\n把可能对症的苗子先==哗啦筛出来==。\n\n常用于找候选分子和预测性质，\n能缩短早期研发，\n但最终仍要实验验证。", "relationsNarrative": "AI for Science\n它是 AI for Science 在生物医药里的典型应用。\n\nDeep Learning\n很多分子筛选与性质预测都由深度学习驱动。\n\nHuman-in-the-loop\nAI 提候选方向，真正定夺还得靠实验专家。", "relations": { "ai-for-science": { "label": "属于…落地场景", "note": "它是科学研究里最热应用之一。" }, "deep-learning": { "label": "常用…建模", "note": "很多分子预测任务靠深度学习。" }, "human-in-the-loop": { "label": "需要…把关", "note": "AI 给建议，实验专家负责验证。" } } } } }, { "id": "ai-election-safeguards", "name": "Election Guard", "layer": "L6", "era": "2024", "publishedAt": "2026-05-31T00:57:33.088Z", "relations": [ { "to": "deepfake" }, { "to": "content-provenance" }, { "to": "ai-regulation" }, { "to": "voice-cloning" } ], "track": "today", "i18n": { "en": { "fullName": "AI Election Safeguards", "factExplain": "A set of rules and tools to reduce AI tricks in elections.", "humanExplain": "Election Guard is a spam filter for election season. It catches the fake “Mayor said WHAT?” post before Uncle Dave forwards it.\n\nIt blocks fake videos and fake voices. You meet it wherever election posts get made or checked.", "humanExplainDisplay": "Election Guard is a ==spam filter==\nfor election season.\nIt catches the ==fake “Mayor said WHAT?” post==\nbefore Uncle Dave forwards it.\n\nIt blocks fake videos and fake voices.\nYou meet it wherever election posts\nget made or checked.", "relationsNarrative": "Deepfake\nElection Guard tries to stop deepfakes from fooling voters.\n\nContent provenance\nContent provenance helps people see if campaign material was changed or faked.\n\nAI-regulation\nAI-regulation turns many election safeguards into laws and platform rules.\n\nVoice cloning\nVoice cloning can copy a candidate or official, so Election Guard watches it closely.", "relations": { "deepfake": { "label": "guards against … fakes", "note": "Deepfakes can show candidates saying things they never said." }, "content-provenance": { "label": "checks sources with …", "note": "Content provenance adds source clues for checking election posts." }, "ai-regulation": { "label": "runs on … rules", "note": "Laws and platform rules make many safeguards real." }, "voice-cloning": { "label": "limits … impersonation", "note": "Voice cloning can fake a candidate's voice in robocalls." } } }, "zh": { "fullName": "AI 选举防护措施", "factExplain": "用于降低 AI 干扰选举风险的一组治理与技术措施。", "humanExplain": "AI 选举防护像给投票季装上小区门禁，谁来塞传单、谁冒充业主，都得先验验。\n\n它常用于识别深伪和消息溯源，帮选民少被假内容带跑。", "humanExplainDisplay": "AI 选举防护像\n==给投票季装上小区门禁==，\n谁来塞传单、谁冒充业主，\n都得先验验。\n\n它常用于识别深伪和消息溯源，\n帮选民少被假内容带跑。", "relationsNarrative": "Deepfake\n选举防护的重要目标之一，是压制伪造内容误导选民。\n\nContent provenance\n内容来源标记能帮助判断竞选材料是否被篡改或伪造。\n\nAI regulation\n很多选举防护措施最终要通过法律和平台规则落地。\n\nVoice-cloning\n语音克隆会冒充候选人或官员，属于重点防范对象。", "relations": { "deepfake": { "label": "防范…造假", "note": "重点应对伪造音视频误导选民。" }, "content-provenance": { "label": "用…验来源", "note": "给内容加出处线索，便于核验真伪。" }, "ai-regulation": { "label": "落到…规则", "note": "很多措施需靠法律与平台制度执行。" }, "voice-cloning": { "label": "限制…冒充", "note": "防止伪造候选人声音进行操纵传播。" } } } } }, { "id": "ai-enabled-terrorism", "name": "AI-enabled terrorism", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "ai-influence-operation" }, { "to": "deepfake" }, { "to": "ai-biosecurity" }, { "to": "frontier-model-access-control" } ], "track": "today", "seo": { "zh": { "title": "AI 赋能恐怖主义 是什么?坏人的群发喇叭,一文看懂 — AI Rookies", "description": "利用 AI 放大恐怖活动能力的风险。人话版一分钟讲明白。" }, "en": { "title": "What Is AI-enabled terrorism? Explained in Plain English", "description": "The risk that AI helps terrorists spread harm faster and farther. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "AI-enabled terrorism", "factExplain": "The risk that AI helps terrorists spread harm faster and farther.", "humanExplain": "AI-enabled terrorism is like a bully getting the school loudspeaker. The hate is not new. The megaphone just got bigger.\n\nIt can boost propaganda. It can also make recruiting and coordination easier. You meet this idea in security policy and platform safety work.", "humanExplainDisplay": "AI-enabled terrorism is like a bully getting the ==school loudspeaker==.\nThe hate is not new.\nThe ==megaphone== just got bigger.\n\nIt can boost propaganda.\nIt can also make recruiting and coordination easier.\nYou meet this idea in security policy\nand platform safety work.", "relationsNarrative": "AI Influence Operations\nIt can scale propaganda, recruiting, and incitement.\n\nDeepfake\nDeepfakes can fake videos and spread panic.\n\nAI biosecurity\nBiosecurity is one of its most sensitive abuse areas.\n\nFrontier Model Access Control\nAccess controls on powerful models can lower the chance of abuse.", "relations": { "ai-influence-operation": { "label": "spreads incitement through …", "note": "AI can make propaganda and incitement cheaper to spread." }, "deepfake": { "label": "fakes content with …", "note": "Fake videos can fuel panic and mislead people." }, "ai-biosecurity": { "label": "raises … risks", "note": "Biosecurity is a very sensitive abuse risk." }, "frontier-model-access-control": { "label": "needs … limits", "note": "Access control can lower the chance of high-risk abuse." } } }, "zh": { "fullName": "AI 赋能恐怖主义", "factExplain": "利用 AI 放大恐怖活动能力的风险。", "humanExplain": "AI 赋能恐怖主义，像坏人捡到群发喇叭：恶意没升级，煽动和恐慌却跑得飞快。\n\n它放大宣传、招募和协调风险，常用于安全治理与平台风控。", "humanExplainDisplay": "AI 赋能恐怖主义，\n像坏人捡到==群发喇叭==：\n恶意没升级，\n煽动和恐慌却跑得飞快。\n\n它放大宣传、招募\n和协调风险，\n常用于安全治理与平台风控。", "relationsNarrative": "AI Influence Operations\n它可把宣传、招募和煽动规模化。\n\nDeepfake\nDeepfake 可被用于伪造影像与煽动传播。\n\nAI Biosecurity\n生物安全是其最敏感的滥用方向之一。\n\nFrontier Model Access Control\n高能力模型访问控制能降低滥用概率。", "relations": { "ai-influence-operation": { "label": "借…扩散煽动", "note": "宣传煽动可被更低成本放大。" }, "deepfake": { "label": "借…伪造内容", "note": "伪造影像会加剧恐慌与误导。" }, "ai-biosecurity": { "label": "涉及…风险", "note": "生物安全是高敏感滥用方向。" }, "frontier-model-access-control": { "label": "需要…约束", "note": "访问控制可降低高危滥用概率。" } } } } }, { "id": "ai-export-controls", "name": "AI Export Controls", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "ai-chip" }, { "to": "compute-race" }, { "to": "sovereign-ai" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "AI Export Controls", "factExplain": "Government rules limit AI chips, models, or tech from crossing borders.", "humanExplain": "Think airport security, but for AI. Even a superstar chip needs a stamp.\n\nThese rules affect who gets AI chips and models. They also shape where companies build data centers and sell AI services.", "humanExplainDisplay": "Think ==airport security==,\nbut for AI.\nEven a ==superstar chip== needs a stamp.\n\nThese rules affect who gets AI chips and models.\nThey also shape where companies build data centers\nand sell AI services.", "relationsNarrative": "AI chip\nExport controls often limit where high-end AI chips can go.\n\nCompute-race\nThey can speed up or slow down each country's compute race.\n\nSovereign AI\nForeign limits push countries to build local AI and compute.\n\nAI-regulation\nAI export controls are a national security tool in AI-regulation.", "relations": { "ai-chip": { "label": "limits export of …", "note": "High-end AI chips are often the main target." }, "compute-race": { "label": "shapes the pace of …", "note": "Controls change the speed of the global compute race." }, "sovereign-ai": { "label": "pushes countries toward …", "note": "Limits from abroad make local models and compute more attractive." }, "ai-regulation": { "label": "belongs in the … toolbox", "note": "It is a national security tool for AI policy." } } }, "zh": { "fullName": "AI 出口管制", "factExplain": "政府限制 AI 芯片、模型或相关技术跨境流动的政策。", "humanExplain": "这不是普通寄快递，是给芯片和模型办“出境批文”：东西再抢手，没章也别想过关。\n\n会影响芯片供应、模型获取和跨境部署，也会改变企业建厂与卖服务的路线。", "humanExplainDisplay": "这不是普通寄快递，\n是给芯片和模型办\n==出境批文==：\n没章也别想==过关==。\n\n会影响芯片供应、\n模型获取和跨境部署；\n也会改变企业建厂\n与卖服务路线。", "relationsNarrative": "AI chip\n出口管制常直接限制高端 AI 芯片流向。\n\nCompute-race\n它会改变各国获取算力的速度与差距。\n\nSovereign AI\n外部受限时，各国更会推动本土 AI 自主化。\n\nAI-regulation\n它是 AI 治理里偏国家安全的一类政策工具。", "relations": { "ai-chip": { "label": "限制…出口", "note": "高端 AI 芯片常是管制重点。" }, "compute-race": { "label": "左右…节奏", "note": "管制会改变各国算力追赶速度。" }, "sovereign-ai": { "label": "推动…建设", "note": "受限后更想自建本土模型与算力。" }, "ai-regulation": { "label": "属于…工具箱", "note": "它是国家层面的产业与安全手段。" } } } } }, { "id": "ai-file-permissions", "name": "AI File Permissions", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "ai-file-upload" }, { "to": "agent-security" }, { "to": "data-privacy" }, { "to": "permission-fatigue" } ], "track": "today", "i18n": { "en": { "fullName": "AI File Permissions", "factExplain": "Rules for the files AI may open, change, or share.", "humanExplain": "AI file permissions are a hall pass for your computer. The AI can open the homework folder, but your diary stays locked.\n\nThey stop the AI from touching the wrong files. You meet them in file uploads and office agents.", "humanExplainDisplay": "AI file permissions are a ==hall pass==\nfor your computer.\nThe AI can open the homework folder,\nbut your ==diary stays locked==.\n\nThey stop the AI\nfrom touching the wrong files.\nYou meet them in file uploads\nand office agents.", "relationsNarrative": "AI File Upload\nFile upload gives the AI material. File permissions decide how much it may use.\n\nAgent Security\nFile permissions keep an agent in the folders it should use.\n\nData-privacy\nThey lower the chance of private files being read or sent out.\n\nPermission fatigue\nToo many permission prompts can make people click Allow with their eyes closed.", "relations": { "ai-file-upload": { "label": "limits what … can use", "note": "After upload, permissions decide where the AI may look." }, "agent-security": { "label": "supports safe access for …", "note": "File permissions are a basic safety fence for agents." }, "data-privacy": { "label": "protects … from leaks", "note": "Fewer file rights mean fewer ways to leak data." }, "permission-fatigue": { "label": "can cause …", "note": "Too many pop-ups make people click Allow without thinking." } } }, "zh": { "fullName": "AI 文件权限", "factExplain": "控制 AI 可读写或共享哪些文件的权限机制。", "humanExplain": "AI 文件权限像校园借书证：课本随便翻，老师答案册想顺走，门儿都没有。\n\n用于上传文件和办公智能体，防误读、误删、外泄。", "humanExplainDisplay": "AI 文件权限像==校园借书证==：\n课本随便翻，\n老师答案册想顺走，\n==门儿都没有==。\n\n用于上传文件和办公智能体，\n防误读、误删、外泄。", "relationsNarrative": "AI File Upload\n文件上传解决给资料，文件权限决定能用多少。\n\nAgent Security\n文件权限把智能体关进该去的房间。\n\nData-privacy\n它减少敏感文件被误读或外传的风险。\n\nPermission Fatigue\n权限提示太频繁，用户可能闭眼点同意。", "relations": { "ai-file-upload": { "label": "限制…的可用范围", "note": "上传文件后，还要管 AI 能看哪里。" }, "agent-security": { "label": "支撑…的最小权限", "note": "文件权限是智能体安全的基础护栏。" }, "data-privacy": { "label": "保护…不外泄", "note": "少给文件权限，就少一条泄露路径。" }, "permission-fatigue": { "label": "可能引发…", "note": "权限弹窗太多，用户会直接乱点。" } } } } }, { "id": "ai-file-upload", "name": "AI File Upload", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "document-parsing" }, { "to": "ocr" }, { "to": "context-window" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "AI File Upload", "factExplain": "A way to give local files to AI for reading and processing.", "humanExplain": "It is like dropping a fat folder on a homework helper’s desk. You stop typing page by page, and the AI starts reading.\n\nYou use it for PDFs, sheets, or screenshots. The AI can summarize them, answer questions, or pull out details.", "humanExplainDisplay": "It is like dropping a ==fat folder==\non a homework helper’s desk.\nYou stop typing page by page,\nand the AI ==starts reading==.\n\nYou use it for PDFs, sheets, or screenshots.\nThe AI can summarize them,\nanswer questions,\nor pull out details.", "relationsNarrative": "Document parsing\nFile upload often uses Document parsing to pull out the layout and text.\n\nOCR\nOCR reads the words in scans, photos, or screenshots first.\n\nContext-window\nContext-window limits how much the model can read at once.\n\nData-privacy\nData-privacy risk appears fast with contracts, medical records, or private files.", "relations": { "document-parsing": { "label": "relies on … to read files", "note": "Document parsing turns the file into content the model can use." }, "ocr": { "label": "uses … to read image text", "note": "Scans and screenshots often need OCR first." }, "context-window": { "label": "is limited by … size", "note": "A very long file may not fit into one input." }, "data-privacy": { "label": "raises … risk", "note": "Think before uploading private or sensitive files." } } }, "zh": { "fullName": "AI 文件上传", "factExplain": "把本地文件交给 AI 读取和处理的入口。", "humanExplain": "跟把体检单往医院窗口一递差不多：你不用逐项念，AI 直接==看单子开工==，省得==嘴皮子打字一起累==。\n\n常用于上传 PDF、表格、截图，让 AI 做总结、问答和信息提取。", "humanExplainDisplay": "跟把体检单往医院窗口一递差不多：\n你不用逐项念，AI 直接==看单子开工==，\n省得==嘴皮子打字一起累==。\n\n常用于上传 PDF、表格、\n截图，\n让 AI 做总结、问答和信息提取。", "relationsNarrative": "Document Parsing\n文件上传后，常先靠它把版面和内容拆出来。\n\nOCR\n遇到扫描件、照片或截图时，常要先识别文字。\n\nContext Window\n文件再长，最后也得受模型一次能读多少限制。\n\nData Privacy\n上传合同、病历这类资料时，隐私风险会立刻出现。", "relations": { "document-parsing": { "label": "依赖…读文件", "note": "先把文件拆成模型能理解的内容。" }, "ocr": { "label": "用…识别图片字", "note": "扫描件和截图常要先识字。" }, "context-window": { "label": "受…限制大小", "note": "文件太长时，装不进一次输入。" }, "data-privacy": { "label": "牵涉…风险", "note": "上传前要想清楚资料是否敏感。" } } } } }, { "id": "ai-financial-advisor", "name": "Robo-advisor", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "human-in-the-loop" }, { "to": "ai-bias" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "Robo-advisor", "factExplain": "A digital advisor that uses AI to give investing and money advice.", "humanExplain": "A robo-advisor is like a bank desk open at 2 a.m. It does the math fast, but may miss your dream of a tiny beach shack.\n\nYou meet it in investing apps. It can suggest a money mix, but risky calls need a human check.", "humanExplainDisplay": "A robo-advisor is like a ==bank desk==\n==open at 2 a.m.==\nIt does the math fast,\nbut may miss your dream\nof a tiny beach shack.\n\nYou meet it in investing apps.\nIt can suggest a money mix,\nbut risky calls need a human check.", "relationsNarrative": "Agent\nA robo-advisor may use an Agent to analyze and run the steps.\n\nHuman-in-the-loop\nRisky money advice usually needs a human final check.\n\nAI-bias\nBiased data or rules can make the advice unfair.\n\nAI-regulation\nMoney advice must follow rules on safety and responsibility.", "relations": { "agent": { "label": "often works as …", "note": "Many robo-advisors use tools and run steps like an Agent." }, "human-in-the-loop": { "label": "needs … review", "note": "Risky money advice usually needs a human final check." }, "ai-bias": { "label": "can be affected by …", "note": "Biased data can make advice fit some people badly." }, "ai-regulation": { "label": "is limited by …", "note": "Money advice must follow rules on safety and responsibility." } } }, "zh": { "fullName": "AI 理财顾问", "factExplain": "用 AI 提供投资与财务建议的数字顾问。", "humanExplain": "智能投顾像手机里的理财经理，不陪你喝茶，只按风险口味配菜。\n\n常用于基金组合和养老金配置，门槛低；但不保证稳赚。", "humanExplainDisplay": "智能投顾像==手机里的理财经理==，\n不陪你喝茶，\n只==按风险口味配菜==。\n\n常用于基金组合和养老金配置，\n门槛低；\n但不保证稳赚。", "relationsNarrative": "Agent\nAI 理财顾问常被做成 Agent 来分析并执行流程。\n\nHuman-in-the-loop\n涉及高风险财务建议时，通常需要人工复核。\n\nAI-bias\n数据或规则有偏差，建议就可能对人不公平。\n\nAI-regulation\n理财建议牵涉合规、责任与用户保护要求。", "relations": { "agent": { "label": "常做成…形态", "note": "很多 AI 理财顾问会自主调用工具执行流程。" }, "human-in-the-loop": { "label": "需要…复核", "note": "涉及高风险建议时通常要人工最终确认。" }, "ai-bias": { "label": "可能受…影响", "note": "训练数据偏差会让建议对部分人群失真。" }, "ai-regulation": { "label": "受…约束", "note": "理财建议触及合规与责任边界。" } } } } }, { "id": "ai-finops", "name": "AI FinOps", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "cost-aware-ai-ai" }, { "to": "ai-unit-economics-ai" }, { "to": "llmops" }, { "to": "ai-usage-cap" } ], "track": "today", "i18n": { "en": { "fullName": "AI Financial Operations", "factExplain": "A way to manage and cut the cost of using AI.", "humanExplain": "AI FinOps is the smart meter for your AI apps. One team runs the big model like the office AC in July. The bill waves a tiny red flag.\n\nIt tracks use and splits budgets. It also helps teams pick the right model before costs run wild.", "humanExplainDisplay": "AI FinOps is the ==smart meter==\nfor your AI apps.\nOne team runs the big model\nlike the ==office AC in July==.\nThe bill waves a tiny red flag.\n\nIt tracks use and splits budgets.\nIt also helps teams pick the right model\nbefore costs run wild.", "relationsNarrative": "Cost-aware AI\nAI FinOps turns cost-aware AI into budgets, tracking, and better choices.\n\nAI Unit Economics\nAI FinOps uses task cost to see whether an AI service is worth it.\n\nLLMOps\nAI FinOps often lives inside LLMOps to track use and cost.\n\nUsage cap\nA usage cap is a common way AI FinOps keeps spending under control.", "relations": { "cost-aware-ai-ai": { "label": "puts … into practice", "note": "It turns saving ideas into daily cost control." }, "ai-unit-economics-ai": { "label": "checks … math", "note": "It shows whether each AI call makes money." }, "llmops": { "label": "plugs into … workflows", "note": "It tracks use and cost during model operations." }, "ai-usage-cap": { "label": "sets … guardrails", "note": "Spending caps stop surprise bills from blowing up." } } }, "zh": { "fullName": "AI 财务运营", "factExplain": "管理并优化 AI 使用成本的运营方法。", "humanExplain": "AI FinOps 像给模型接上电费表：谁把大模型当空调猛开，账单马上亮灯。\n\n用于监控用量、分摊预算和选模型，防止成本失控。", "humanExplainDisplay": "AI FinOps 像给模型\n接上==电费表==：\n谁把大模型当空调猛开，\n账单马上==亮灯==。\n\n用于监控用量、分摊预算和选模型，\n防止成本失控。", "relationsNarrative": "Cost-aware AI\nAI FinOps 把成本意识落到预算、监控和优化。\n\nAI Unit Economics\n它用单次任务成本判断 AI 服务是否划算。\n\nLLMOps\n它常嵌入模型部署运维，跟踪用量与费用。\n\nUsage cap\n额度上限是它控制成本失控的常用手段。", "relations": { "cost-aware-ai-ai": { "label": "落实…策略", "note": "把省钱原则变成日常成本管理。" }, "ai-unit-economics-ai": { "label": "核算…账本", "note": "看清每次调用到底赚不赚钱。" }, "llmops": { "label": "接入…流程", "note": "在部署运维中追踪用量和费用。" }, "ai-usage-cap": { "label": "设置…防线", "note": "用额度限制避免账单突然爆雷。" } } } } }, { "id": "ai-for-consumers", "name": "AI for Consumers", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "personal-ai-apps" }, { "to": "ai-super-app" }, { "to": "ai-monetization" }, { "to": "ai-adoption-curve" } ], "track": "today", "i18n": { "en": { "fullName": "AI for Consumers", "factExplain": "AI moving from expert tools into everyday apps people use at home.", "humanExplain": "AI used to be the cafe espresso machine. Now it is the microwave button you press in socks.\n\nYou meet it in chat apps, photo tools, and shopping. The real test is simple: will normal people open it every day?", "humanExplainDisplay": "AI used to be the ==cafe espresso machine==.\nNow it is the ==microwave button==\nyou press in socks.\n\nYou meet it in chat apps,\nphoto tools, and shopping.\nThe real test is simple:\nwill normal people open it every day?", "relationsNarrative": "Personal AI apps\nAI for Consumers first shows up in apps people open every day.\n\nAI super app\nDaily needs can push one AI app to become the main doorway.\n\nAI monetization\nAI for Consumers only works if people pay and come back.\n\nAdoption Curve\nAI for Consumers moves AI from early fans to everyday users.", "relations": { "personal-ai-apps": { "label": "shows up in …", "note": "Consumer AI often starts as personal apps people use every day." }, "ai-super-app": { "label": "pushes … to form", "note": "Daily needs can turn one app into the main AI doorway." }, "ai-monetization": { "label": "tests …", "note": "People must keep paying for consumer AI to last." }, "ai-adoption-curve": { "label": "speeds up …", "note": "It moves AI from early fans to the mass market." } } }, "zh": { "fullName": "AI 消费化 / AI+消费", "factExplain": "AI 从专业工具走向大众日常消费场景的过程。", "humanExplain": "原先像药柜顶层的处方药，如今成了家里抽屉常备药：不用懂配方，顺手就敢拿来用。\n\n常见于聊天、修图和购物，重点是普通人是否愿意天天打开。", "humanExplainDisplay": "原先像药柜顶层的\n==处方药==，\n如今成了家里抽屉\n常备药：\n不用懂配方，\n顺手就敢==拿来用==。\n\n常见于聊天、修图\n和购物，\n重点是普通人是否\n愿意天天打开。", "relationsNarrative": "Personal AI apps\n消费化最先体现在个人用户天天打开的应用里。\n\nAI super app\n高频、多场景需求会推动超级入口出现。\n\nAI monetization\n消费化能不能跑通，最终要看付费与复购。\n\nAdoption Curve\n它对应 AI 从少数尝鲜者走向大众用户。", "relations": { "personal-ai-apps": { "label": "落到…里", "note": "消费化通常先表现为个人应用爆发。" }, "ai-super-app": { "label": "推动…形成", "note": "高频消费需求会催生超级入口。" }, "ai-monetization": { "label": "决定…空间", "note": "能否持续付费是消费化关键考题。" }, "ai-adoption-curve": { "label": "加速…扩散", "note": "从早期玩家走向大众市场。" } } } } }, { "id": "ai-for-science", "name": "AI for Science", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "ai-assisted-research" }, { "to": "ai-math-discovery" }, { "to": "multimodal" }, { "to": "foundation-model" } ], "track": "today", "i18n": { "en": { "fullName": "AI for Science", "factExplain": "Using AI to help scientists study data and find new discoveries.", "humanExplain": "AI for Science is like a lab partner with endless snacks. When the lights go off, it is still chasing clues.\n\nIt helps scientists read papers and scan data faster. It helps in drug and materials research. It also helps in climate and math.", "humanExplainDisplay": "AI for Science is like\n==a lab partner with endless snacks==.\nWhen the lights go off,\nit is still ==chasing clues==.\n\nIt helps scientists read papers\nand scan data faster.\nIt helps in drug and materials research.\nIt also helps in climate and math.", "relationsNarrative": "AI-assisted Research\nAI for Science is the wider area behind AI-assisted Research.\n\nAI Math Discovery\nAI Math Discovery is a key use of AI for Science in basic science.\n\nMultimodal AI\nAI for Science often uses Multimodal AI because lab data comes in many forms.\n\nFoundation-model\nMany AI for Science tools build on a Foundation-model.", "relations": { "ai-assisted-research": { "label": "lands in …", "note": "AI-assisted Research is AI for Science at work in real research tasks." }, "ai-math-discovery": { "label": "extends into …", "note": "AI Math Discovery is one key branch of AI for Science." }, "multimodal": { "label": "often uses …", "note": "Science data can be words, images, tables, or lab results." }, "foundation-model": { "label": "often builds on …", "note": "Many science tools start with a broad foundation model." } } }, "zh": { "fullName": "AI for Science（AI 用于科学研究）", "factExplain": "用 AI 辅助科学研究与发现的方向。", "humanExplain": "像给实验室请了个卷王师弟：文献先读、数据先筛、可疑线索先跑，导师睡了它还在干。\n\n常用于药物、材料、气候和数学研究，帮人更快找规律和筛方案。", "humanExplainDisplay": "像给实验室请了个==卷王师弟==：\n文献先读、数据先筛、\n可疑线索先==跑==，\n导师睡了它还在干。\n\n常用于药物、材料、\n气候和数学研究，\n帮人更快找规律和筛方案。", "relationsNarrative": "AI-assisted Research\n它是 AI 辅助研究的更大范畴，覆盖更多科学学科。\n\nAI Math Discovery\n数学发现是它在基础科学中的典型应用方向。\n\nMultimodal\n科研常要同时理解文本、图像、表格等多种数据。\n\nFoundation-model\n许多科研系统建立在通用基础模型能力之上。", "relations": { "ai-assisted-research": { "label": "落到…场景", "note": "它在科研场景中的直接应用形态。" }, "ai-math-discovery": { "label": "延伸到…方向", "note": "数学发现是其代表性分支之一。" }, "multimodal": { "label": "常结合…能力", "note": "科研数据常不止文字，还含图表和实验结果。" }, "foundation-model": { "label": "常建立在…上", "note": "很多科研应用以通用大模型为底座。" } } } } }, { "id": "ai-future-scenarios", "name": "AI Future Scenarios", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "agi" }, { "to": "alignment" }, { "to": "ai-regulation" }, { "to": "ai-anxiety" } ], "track": "today", "seo": { "zh": { "title": "AI 未来情景 是什么?未来先摆几盘棋,一文看懂 — AI Rookies", "description": "用多种假设描绘 AI 可能未来的分析方法。人话版一分钟讲明白。" }, "en": { "title": "What Is AI Future Scenarios? Picnic Plans for AI Futures", "description": "A way to map possible AI futures using several clear “what if” stories. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "AI Future Scenarios", "factExplain": "A way to map possible AI futures using several clear “what if” stories.", "humanExplain": "AI future scenarios are not a magic 8-ball. They are like planning a picnic with sunny, rainy, and “the ants invade” plans.\n\nPeople use them to discuss AGI, rules, and jobs. They stop one scary future from filling the whole room.", "humanExplainDisplay": "AI future scenarios are not a ==magic 8-ball==.\nThey are like planning a picnic\nwith sunny, rainy,\nand ==“the ants invade”== plans.\n\nPeople use them to discuss AGI,\nrules, and jobs.\nThey stop one scary future\nfrom filling the whole room.", "relationsNarrative": "AGI\nAI future scenarios often use AGI as a key fork in the road.\n\nAlignment\nGood or bad alignment changes where many scenarios go.\n\nAI-regulation\nStronger or weaker rules change the pace of risk and progress.\n\nAI-anxiety\nSeeing several paths can calm fear of one fixed future.", "relations": { "agi": { "label": "often centers on …", "note": "Many scenarios ask when AGI might arrive." }, "alignment": { "label": "depends on …", "note": "Good or bad alignment can change the future path." }, "ai-regulation": { "label": "connects to …", "note": "Scenarios turn vague risks into policy choices." }, "ai-anxiety": { "label": "eases …", "note": "Several storylines can break one big fear into smaller fears." } } }, "zh": { "fullName": "AI 未来情景", "factExplain": "用多种假设描绘 AI 可能未来的分析方法。", "humanExplain": "AI 未来情景不是押宝猜彩票，而是下棋先摆几盘变化：每条路都先演一遍。\n\n用于 AGI、监管和就业讨论，帮人跳出单一路线。", "humanExplainDisplay": "AI 未来情景不是\n==押宝猜彩票==，\n而是下棋先摆几盘变化：\n每条路都先演一遍。\n\n用于 AGI、监管\n和就业讨论，\n帮人跳出单一路线。", "relationsNarrative": "AGI\n它常把 AGI 作为关键分叉点。\n\nAlignment\n对齐成败决定许多情景的走向。\n\nAI-regulation\n监管力度会改变风险与创新节奏。\n\nAI-anxiety\n多情景视角能缓解单一路线恐惧。", "relations": { "agi": { "label": "围绕…展开", "note": "很多情景围绕 AGI 何时到来。" }, "alignment": { "label": "衡量…影响", "note": "对齐好坏会改变未来分叉。" }, "ai-regulation": { "label": "连接…议题", "note": "情景能把抽象风险变成政策题。" }, "ai-anxiety": { "label": "缓解…", "note": "多种剧本能拆掉单一恐惧。" } } } } }, { "id": "ai-generated-advertising", "name": "AI-generated Advertising", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "prompt-engineering" }, { "to": "ai-monetization" }, { "to": "content-provenance" } ], "track": "today", "i18n": { "en": { "fullName": "AI-generated Advertising", "factExplain": "Generative AI that makes or rewrites ad text, images, or videos.", "humanExplain": "It is like a coffee-fueled ad intern who never clocks out. Ask for a sneaker banner, and five versions land before your toast pops.\n\nTeams use it to make many ad versions for different people. It saves time, but humans still need to check each ad.", "humanExplainDisplay": "It is like a ==coffee-fueled ad intern==\nwho never clocks out.\nAsk for a ==sneaker banner==,\nand five versions land\nbefore your toast pops.\n\nTeams use it to make many ad versions\nfor different people.\nIt saves time,\nbut humans still need to check each ad.", "relationsNarrative": "Generative Model\nAI-generated ads use a generative model to make text, images, and videos.\n\nPrompt-engineering\nPrompt-engineering tells the ad what to sell and how to sound.\n\nAI monetization\nAI-generated ads are a common way AI helps businesses earn money.\n\nContent provenance\nContent provenance helps people see if an ad was made by AI.", "relations": { "generative-model": { "label": "uses … to make ads", "note": "It makes the ad text, images, and videos." }, "prompt-engineering": { "label": "steers style with …", "note": "Good prompts set the selling point and tone." }, "ai-monetization": { "label": "supports …", "note": "Ads are a direct way AI helps businesses earn money." }, "content-provenance": { "label": "needs …", "note": "Source labels show when an ad was made by AI." } } }, "zh": { "fullName": "AI 生成广告", "factExplain": "用生成式 AI 自动制作或改写广告素材。", "humanExplain": "AI 生成广告像凌晨还在线的乙方：横幅、文案、短视频，催完就出稿。\n\n用于批量投放和个性化测试，省时间，也更要审核。", "humanExplainDisplay": "AI 生成广告像\n==凌晨还在线的乙方==：\n横幅、文案、短视频，\n==催完就出稿==。\n\n用于批量投放\n和个性化测试，\n省时间，也更要审核。", "relationsNarrative": "Generative Model\n生成式模型负责产出广告文案、图片和视频素材。\n\nPrompt-engineering\n提示词决定广告卖点、语气和目标人群是否对味。\n\nAI Monetization\n广告是 AI 帮企业降本增效并变现的常见场景。\n\nContent Provenance\n来源标记能帮助用户识别广告是否由 AI 生成。", "relations": { "generative-model": { "label": "用…生成素材", "note": "广告文案、图片、视频都靠它产出。" }, "prompt-engineering": { "label": "靠…控制风格", "note": "好提示词决定卖点和语气是否对味。" }, "ai-monetization": { "label": "服务…变现", "note": "广告是 AI 商业化最直接的入口之一。" }, "content-provenance": { "label": "需要…标记来源", "note": "生成广告越多，来源标记越重要。" } } } } }, { "id": "ai-generated-evidence", "name": "AI-generated evidence", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "deepfake" }, { "to": "content-provenance" }, { "to": "ai-regulation" }, { "to": "ai-election-safeguards" } ], "track": "today", "i18n": { "en": { "fullName": "AI-generated evidence", "factExplain": "Content made by AI and used as proof.", "humanExplain": "AI evidence is like a fake hall pass signed by the principal. It looks official, but the principal never touched it.\n\nYou may meet it in court or elections. It often needs a source check before anyone trusts it.", "humanExplainDisplay": "AI evidence is like a ==fake hall pass==\nsigned by the principal.\nIt ==looks official==,\nbut the principal never touched it.\n\nYou may meet it in court or elections.\nIt often needs a source check\nbefore anyone trusts it.", "relationsNarrative": "Deepfake\nAI evidence often appears as fake audio or video.\n\nContent provenance\nContent provenance helps check its source and truth.\n\nAI-regulation\nAI evidence pushes evidence rules and platform rules to update.\n\nElection Guard\nIn elections, fake evidence can put safeguards under heavy stress.", "relations": { "deepfake": { "label": "often appears as …", "note": "Fake audio or video is a common form of AI evidence." }, "content-provenance": { "label": "checks source with …", "note": "Provenance helps show where the content came from." }, "ai-regulation": { "label": "pushes … to act", "note": "AI evidence forces rules and platforms to update." }, "ai-election-safeguards": { "label": "tests …", "note": "Fake evidence can shake trust during elections." } } }, "zh": { "fullName": "AI 生成证据", "factExplain": "由 AI 生成并被当作证据使用的内容。", "humanExplain": "像相亲局上有人把聊天截图、语音和合照全修到天衣无缝，还当铁证甩出来。\n\n会直接影响能不能采信，通常得先做溯源核验。", "humanExplainDisplay": "像相亲局上有人把\n聊天截图、语音和合照\n全修到==天衣无缝==，\n还当==铁证==甩出来。\n\n会直接影响\n能不能采信，\n通常得先做\n溯源核验。", "relationsNarrative": "Deepfake\nAI 生成证据常以伪造音视频形式出现。\n\nContent provenance\n内容溯源可帮助核验它的来源与真伪。\n\nAI-regulation\n它会推动证据规则和平台监管更新。\n\nElection Guard\n选举场景里，伪造证据会直接冲击防护机制。", "relations": { "deepfake": { "label": "常以…出现", "note": "伪造音视频是典型表现形式。" }, "content-provenance": { "label": "靠…核验来源", "note": "来源链能帮助判断是否被伪造。" }, "ai-regulation": { "label": "推动…介入", "note": "它倒逼证据规则与平台监管更新。" }, "ai-election-safeguards": { "label": "考验…机制", "note": "选举场景尤其怕伪造证据带节奏。" } } } } }, { "id": "ai-governance-framework", "name": "AI Governance", "layer": "L6", "era": "2023", "publishedAt": "2026-05-31T00:57:33.087Z", "relations": [ { "to": "ai-regulation" }, { "to": "alignment" }, { "to": "human-in-the-loop" }, { "to": "third-party-ai-evaluation" } ], "track": "today", "i18n": { "en": { "fullName": "AI Governance Framework", "factExplain": "A rule framework for how AI is built, launched, and used.", "humanExplain": "AI governance is the house rules on the fridge. So when the robot dog eats the homework, nobody yells, “Whose problem is this?”\n\nCompanies use it before AI goes live. Governments use it for rules and high-risk checks, so responsibility stays clear.", "humanExplainDisplay": "AI governance is the ==house rules== on the fridge.\nSo when the robot dog eats the homework,\nnobody yells,\n==“Whose problem is this?”==\n\nCompanies use it before AI goes live.\nGovernments use it for rules and high-risk checks,\nso responsibility stays clear.", "relationsNarrative": "AI-regulation\nAI governance turns outside regulation into internal rules.\n\nAlignment\nIt checks whether model behavior stays within human goals and limits.\n\nHuman-in-the-loop\nFor high-risk tasks, AI governance often keeps a human as the final decider.\n\nThird-party AI evaluation\nThird-party eval can serve as an outside review step.", "relations": { "ai-regulation": { "label": "turns … into rules", "note": "A governance framework turns outside regulation into internal steps." }, "alignment": { "label": "keeps … on target", "note": "It checks whether model goals stay close to human intent." }, "human-in-the-loop": { "label": "puts … in charge", "note": "High-risk work often keeps a human as the final decider." }, "third-party-ai-evaluation": { "label": "uses … for review", "note": "Independent review helps check risks and rule-following." } } }, "zh": { "fullName": "AI 治理框架", "factExplain": "用于规范 AI 开发、部署与使用的制度框架。", "humanExplain": "它像给外卖骑手画禁行区：能跑得快，但别冲进厨房。\n\n它用于企业制度、监管合规和风险评估，让创新别一路裸奔。", "humanExplainDisplay": "它像==给外卖骑手画禁行区==：\n能跑得快，\n但==别冲进厨房==。\n\n它用于企业制度、\n监管合规和风险评估，\n让创新别一路裸奔。", "relationsNarrative": "AI-regulation\n治理框架常把外部监管要求翻成内部规则。\n\nAlignment\n它关心模型行为是否符合人的目标与边界。\n\nHuman-in-the-loop\n高风险任务里，治理框架常要求人来最终把关。\n\nThird-party AI evaluation\n第三方评估可作为治理框架里的外部审查环节。", "relations": { "ai-regulation": { "label": "落地…要求", "note": "治理框架常把监管要求变成内部流程。" }, "alignment": { "label": "约束…目标", "note": "它要求模型目标与人类意图别跑偏。" }, "human-in-the-loop": { "label": "安排…把关", "note": "高风险场景常要求人类保留最终决定权。" }, "third-party-ai-evaluation": { "label": "借…做审查", "note": "独立评估常用于验证风险与合规性。" } } } } }, { "id": "ai-industrial-policy", "name": "AI Industrial Policy", "layer": "L6", "era": "2017", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "ai-export-controls" }, { "to": "sovereign-ai" }, { "to": "compute-race" }, { "to": "ai-chip" } ], "track": "today", "i18n": { "en": { "fullName": "AI Industrial Policy", "factExplain": "Government plans that steer how a country's AI industry grows.", "humanExplain": "AI industrial policy is a strict coach for the AI team. It picks the field, buys the cleats, and yells, “No giving our best ball away!”\n\nYou meet it in chip funding and data-center plans. It also shapes AI workers and overseas sales.", "humanExplainDisplay": "AI industrial policy is a ==strict coach==\nfor the AI team.\nIt picks the field,\nbuys the cleats,\nand yells,\n“==No giving our best ball away!==”\n\nYou meet it in chip funding\nand data-center plans.\nIt also shapes AI workers\nand overseas sales.", "relationsNarrative": "AI Export Controls\nAI industrial policy uses export controls to stop key tech from leaving.\n\nSovereign AI\nSovereign AI is the goal of staying in control at home.\n\nCompute-race\nAI industrial policy can turn computing power into a national race.\n\nAI chip\nChip support is often its strongest tool.", "relations": { "ai-export-controls": { "label": "sets checkpoints with …", "note": "Export controls are the hard brake inside AI industrial policy." }, "sovereign-ai": { "label": "pushes toward …", "note": "Sovereign AI is the goal of keeping AI under national control." }, "compute-race": { "label": "speeds up …", "note": "Policy turns computing power into a national race." }, "ai-chip": { "label": "backs the … industry", "note": "Chips are often the strongest tool in the policy box." } } }, "zh": { "fullName": "AI Industrial Policy（AI 产业政策）", "factExplain": "政府引导 AI 产业发展的政策组合。", "humanExplain": "AI 产业政策像给高铁铺轨：往哪开、谁上车，哪些货不能出站，全写进信号灯。\n\n用于引导芯片、算力、人才布局，并影响企业出海。", "humanExplainDisplay": "AI 产业政策像==给高铁铺轨==：\n往哪开、谁上车，\n哪些货==不能出站==，\n全写进信号灯。\n\n用于引导芯片、算力、人才布局，\n并影响企业出海。", "relationsNarrative": "AI Export Controls\n出口管制是产业政策限制关键技术外流的工具。\n\nSovereign AI\n主权 AI 是产业政策追求自主可控的目标。\n\nCompute Race\n产业政策会把算力竞争推向国家赛道。\n\nAI Chip\n芯片扶持常是产业政策最硬的抓手。", "relations": { "ai-export-controls": { "label": "设置…卡口", "note": "出口管制是政策里的硬刹车。" }, "sovereign-ai": { "label": "强化…目标", "note": "主权 AI 追求自主可控。" }, "compute-race": { "label": "加速…竞争", "note": "政策把算力推向国家赛道。" }, "ai-chip": { "label": "扶持…产业", "note": "芯片常是最硬的政策抓手。" } } } } }, { "id": "ai-influence-operation", "name": "AI Influence Operations", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "deepfake" }, { "to": "ai-election-safeguards" }, { "to": "content-provenance" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "AI Influence Operations", "factExplain": "Using AI at scale to push public opinion and shape what people believe.", "humanExplain": "AI influence ops are a troll farm with a copy machine. It spits out fake posts, fake photos, and robot voices all day.\n\nIt shows up in public fights and scam ads. In elections, it can make false stories look popular fast.", "humanExplainDisplay": "AI influence ops are a ==troll farm==\nwith a ==copy machine==.\nIt spits out fake posts,\nfake photos,\nand robot voices all day.\n\nIt shows up in public fights and scam ads.\nIn elections,\nit can make false stories look popular fast.", "relationsNarrative": "Deepfake\nAI influence operations often use fake audio or video to look more believable.\n\nElection Guard\nElection Guard helps stop mass opinion attacks during elections.\n\nContent provenance\nContent provenance helps trace where fake content came from.\n\nAI-regulation\nThis risk pushes platforms and governments toward tougher AI-regulation.", "relations": { "deepfake": { "label": "uses … as bait", "note": "Fake audio or video can make the lie feel real." }, "ai-election-safeguards": { "label": "is blocked by …", "note": "Elections need protection from mass opinion attacks." }, "content-provenance": { "label": "is traced by …", "note": "Source labels help track fake content as it spreads." }, "ai-regulation": { "label": "pushes … to step in", "note": "These risks often lead platforms and governments to tighten rules." } } }, "zh": { "fullName": "AI 影响力行动", "factExplain": "用 AI 批量操纵舆论与认知的信息行动。", "humanExplain": "像水军头子开了群控：段子、假图、配音一起刷，热搜和谣言能成排往外冒。\n\n常见于舆论操纵、诈骗宣传和选举干扰，会放大虚假内容扩散。", "humanExplainDisplay": "像水军头子开了==群控==：\n段子、假图、配音一起刷，\n热搜和谣言能\n==成排往外冒==。\n\n常见于舆论操纵、\n诈骗宣传和选举干扰，\n会放大虚假内容扩散。", "relationsNarrative": "Deepfake\n它常借伪造音视频提升内容可信感和传播力。\n\nAI Election Safeguards\n选举防护措施重点防范这类舆论操纵行为。\n\nContent Provenance\n内容来源标记可帮助追踪伪造内容出处。\n\nAI-regulation\n这类风险常推动平台治理与监管规则加码。", "relations": { "deepfake": { "label": "常用…做素材", "note": "伪造音视频能增强迷惑性。" }, "ai-election-safeguards": { "label": "被…重点防范", "note": "选举场景最怕大规模舆论操纵。" }, "content-provenance": { "label": "靠…追来源", "note": "来源标记有助识别伪造传播。" }, "ai-regulation": { "label": "推动…介入", "note": "平台与政府常据此加强治理。" } } } } }, { "id": "ai-inventorship", "name": "AI Inventorship", "layer": "L6", "era": "2019", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "copyright" }, { "to": "ai-regulation" }, { "to": "ai-assisted-research" }, { "to": "ai-drug-discovery" } ], "track": "today", "i18n": { "en": { "fullName": "AI Inventorship", "factExplain": "The legal question of whether AI can be named as a patent inventor.", "humanExplain": "AI Inventorship is a robot winning the school science fair. Whose name goes on the blue ribbon?\n\nIt decides whose name goes on a patent. You meet it in AI-made drugs and new materials.", "humanExplainDisplay": "AI Inventorship is a ==robot winning==\nthe school science fair.\nWhose name goes on\n==the blue ribbon==?\n\nIt decides whose name goes on a patent.\nYou meet it in AI-made drugs\nand new materials.", "relationsNarrative": "Copyright\nBoth ask who should own work made with AI.\n\nAI-regulation\nAI-regulation must say if AI can be named as an inventor.\n\nAI-assisted Research\nMore AI work in research makes inventor credit harder to draw.\n\nAI Drug Discovery\nAI Drug Discovery often pushes this patent question into the spotlight.", "relations": { "copyright": { "label": "shares ownership puzzles with …", "note": "Both ask who owns work made with AI." }, "ai-regulation": { "label": "waits for …", "note": "Rules must say whether AI can be named." }, "ai-assisted-research": { "label": "grows from …", "note": "More AI help makes inventor credit harder." }, "ai-drug-discovery": { "label": "shows up often in …", "note": "Drug patents often bring this question up." } } }, "zh": { "fullName": "AI 发明人资格", "factExplain": "关于 AI 能否成为专利发明人的法律问题。", "humanExplain": "AI 发明人资格像机器人赢了发明大赛：奖状上该写谁的名字？\n\n它决定专利署名和归属，药物、材料研发最常撞上。", "humanExplainDisplay": "AI 发明人资格像==机器人==\n赢了发明大赛：\n奖状上==该写谁==的名字？\n\n它决定专利署名和归属，\n药物、材料研发，\n最常撞上。", "relationsNarrative": "Copyright\n同属 AI 生成成果的权利归属难题。\n\nAI-regulation\n它需要法规明确 AI 能不能署名。\n\nAI-assisted Research\nAI 越深参与研究，发明人越难划线。\n\nAI Drug Discovery\n药物发现常把 AI 发明权问题推到台前。", "relations": { "copyright": { "label": "牵动…边界", "note": "都在问 AI 成果该归谁。" }, "ai-regulation": { "label": "等待…裁定", "note": "发明人资格要靠法规落地。" }, "ai-assisted-research": { "label": "来自…深化", "note": "研究越自动，署名越难分。" }, "ai-drug-discovery": { "label": "常见于…", "note": "药物专利最容易撞上它。" } } } } }, { "id": "ai-ipo", "name": "AI IPO", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "ai-monetization" }, { "to": "ai-adoption-curve" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "AI Initial Public Offering", "factExplain": "An AI company's first public sale of its shares.", "humanExplain": "An AI IPO is Shark Tank with a stock ticker. The shiny demo sits down and shows its bank statements.\n\nIt tests if the AI story is a real business. It can change how excited investors feel about funding AI.", "humanExplainDisplay": "An AI IPO is ==Shark Tank==\nwith a stock ticker.\nThe ==shiny demo== sits down\nand shows its bank statements.\n\nIt tests if the AI story\nis a real business.\nIt can change how excited investors feel\nabout funding AI.", "relationsNarrative": "AI monetization\nAn AI IPO makes investors look harder at revenue and profit.\n\nAdoption Curve\nFaster adoption makes investors more willing to pay for growth.\n\nCompute-race\nThe compute-race shapes the cost story and the company value.", "relations": { "ai-monetization": { "label": "tests … in public", "note": "An IPO pushes an AI company to prove its revenue is real." }, "ai-adoption-curve": { "label": "depends on …", "note": "Faster user adoption gives investors more room to dream." }, "compute-race": { "label": "raises pressure from …", "note": "After an IPO, people ask harder questions about compute costs." } } }, "zh": { "fullName": "AI 公司首次公开募股", "factExplain": "AI 公司第一次向公众发行股票并上市。", "humanExplain": "这一步像相亲见家长：以前朋友圈吹得再神，真上桌了，房本、收入、脾气都得摊开看。\n\n它检验 AI 故事能否落成生意，也会影响市场融资情绪。", "humanExplainDisplay": "这一步像==相亲见家长==：\n以前朋友圈吹得再神，\n真上桌了，房本、收入、脾气\n都得==摊开看==。\n\n它检验 AI 故事能否落成生意，\n也会影响市场融资情绪。", "relationsNarrative": "AI monetization\nAI IPO 会放大市场对收入与利润的审视。\n\nAdoption Curve\n采用速度越快，资本越愿为增长买单。\n\nCompute-race\n算力军备竞赛会直接影响成本与估值叙事。", "relations": { "ai-monetization": { "label": "检验…变现", "note": "上市会逼公司证明收入是否扎实。" }, "ai-adoption-curve": { "label": "受…影响", "note": "用户采用速度会影响市场想象空间。" }, "compute-race": { "label": "放大…压力", "note": "上市后更容易被追问烧钱与投入。" } } } } }, { "id": "ai-kling-ai", "name": "Kling AI", "layer": "L5", "sublayer": "product", "era": "2024", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "multimodal" }, { "to": "seedance" }, { "to": "deepfake" } ], "track": "history", "i18n": { "en": { "fullName": "Kling AI", "factExplain": "An AI product that makes videos from text prompts or images.", "humanExplain": "Kling AI is a tiny film crew in your laptop. Give it a photo or one wild sentence. It shoots a rough trailer before your coffee cools.\n\nPeople use it for ad storyboards. They also use it for short film ideas and visual demos. It works fast, but small details can be hard to steer.", "humanExplainDisplay": "Kling AI is a ==tiny film crew==\nin your laptop.\nGive it a photo\nor ==one wild sentence==.\nIt shoots a rough trailer\nbefore your coffee cools.\n\nPeople use it for ad storyboards.\nThey also use it for short film ideas\nand visual demos.\nIt works fast,\nbut small details can be hard to steer.", "relationsNarrative": "Diffusion\nMany tools like Kling AI use diffusion methods to help make video.\n\nMultimodal AI\nKling AI reads text and images, then creates video.\n\nSeedance\nKling AI and Seedance are both AI video creation tools.\n\nDeepfake\nMore realistic AI video can make fake content easier to spread.", "relations": { "diffusion": { "label": "is often powered by …", "note": "Many video generators use diffusion ideas under the hood." }, "multimodal": { "label": "is a … app", "note": "It works with text, images, and video." }, "seedance": { "label": "shares video generation with …", "note": "Both tools are used to make AI video." }, "deepfake": { "label": "can raise … risks", "note": "Real-looking AI video can be used to fake events." } } }, "zh": { "fullName": "快手的视频生成产品 可灵", "factExplain": "一款可根据文本或图片生成视频的 AI 产品。", "humanExplain": "你甩一句脑洞给导演，它连分镜带运镜一口气拍成样片，灵感还没凉，视频先出来了。\n\n常用于广告分镜、短片创意和视觉 demo，出片快，但细节未必完全可控。", "humanExplainDisplay": "你甩一句==脑洞==给导演，\n它连分镜带运镜\n一口气拍成==样片==，\n灵感还没凉，视频先出来了。\n\n常用于广告分镜、\n短片创意和视觉 demo，\n出片快，但细节未必完全可控。", "relationsNarrative": "Diffusion\n很多这类视频生成产品，底层常建立在扩散方法上。\n\nMultimodal AI\n它同时理解文字、图片并输出视频，属于多模态应用。\n\nSeedance\n两者都主打生成式视频，可放在同类创作工具里理解。\n\nDeepfake\n视频越逼真，被用于伪造内容和误导传播的风险也越高。", "relations": { "diffusion": { "label": "常由…驱动", "note": "很多视频生成能力建立在扩散思路上。" }, "multimodal": { "label": "属于…应用", "note": "它同时处理文字、图片与视频内容。" }, "seedance": { "label": "同属视频生成", "note": "两者都面向生成式视频创作场景。" }, "deepfake": { "label": "也会带来…风险", "note": "生成逼真视频也可能被拿去伪造内容。" } } } } }, { "id": "ai-literacy-ai", "name": "AI Literacy", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "ai-anxiety" }, { "to": "hallucination" }, { "to": "prompt-engineering" }, { "to": "ai-usage-gap" } ], "track": "today", "i18n": { "en": { "fullName": "AI Literacy", "factExplain": "The basic skill to understand, use, and judge AI.", "humanExplain": "AI literacy is like knowing a microwave. You press the right buttons and keep forks out.\n\nIt helps you ask better questions and choose tools. It also helps you check answers at work, in class, and online.", "humanExplainDisplay": "AI literacy is like knowing a ==microwave==.\nYou press the right buttons\nand ==keep forks out==.\n\nIt helps you ask better questions\nand choose tools.\nIt also helps you check answers\nat work, in class, and online.", "relationsNarrative": "AI-anxiety\nAI literacy shows the limits, so the fear gets smaller.\n\nHallucination\nAI literacy helps you know AI can make things up.\n\nPrompt-engineering\nPrompt-engineering is one clear way AI literacy shows.\n\nAI usage gap\nThe same tool works better for people with more AI literacy.", "relations": { "ai-anxiety": { "label": "eases …", "note": "A little know-how beats blind panic." }, "hallucination": { "label": "helps spot …", "note": "When you know AI can make things up, you check it." }, "prompt-engineering": { "label": "shows in …", "note": "Good questions are a clear sign of AI literacy." }, "ai-usage-gap": { "label": "narrows …", "note": "Skill with AI changes how useful the same tool feels." } } }, "zh": { "fullName": "AI 素养", "factExplain": "理解、使用并判断 AI 的基础能力。", "humanExplain": "AI 素养有点像网购老手：会搜、会比、会看追评，也知道哪些“买家秀”一眼就是修过头。\n\n它影响提问、选工具和判断结果，常用在办公、学习与信息辨别。", "humanExplainDisplay": "AI 素养有点像==网购老手==：\n会搜、会比、会看追评，\n也知道哪些“买家秀”\n一眼就是==修过头==。\n\n它影响提问、选工具\n和判断结果，\n常用在办公、学习\n与信息辨别。", "relationsNarrative": "Ai-anxiety\n懂一点它的边界和用法，能少很多无谓焦虑。\n\nHallucination\nAI 素养包括识别模型会胡编，不轻信输出。\n\nPrompt-engineering\n会不会提问，是 AI 素养最直观的体现之一。\n\nAi-usage-gap\n同样有工具，素养差异会拉开使用效果。", "relations": { "ai-anxiety": { "label": "缓解…带来的慌", "note": "懂一点原理，比盲目焦虑更有用。" }, "hallucination": { "label": "帮人识别…", "note": "知道它会瞎编，才不会照单全收。" }, "prompt-engineering": { "label": "体现在会用…", "note": "会提问的人，通常更会把它用顺。" }, "ai-usage-gap": { "label": "缩小…差距", "note": "会不会用，直接拉开使用效果。" } } } } }, { "id": "ai-live-commentator", "name": "AI Live Commentator", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "streaming-multimodal-model" }, { "to": "multimodal" }, { "to": "tts" } ], "track": "today", "i18n": { "en": { "fullName": "AI Live Commentator", "factExplain": "An AI app that watches a live stream and talks about it in real time.", "humanExplain": "It is the friend on your couch during the game. The ball moves, and the nacho expert is already yelling what happened.\n\nYou meet it in sports streams, esports streams, or shopping streams. It covers quiet gaps and keeps the show lively.", "humanExplainDisplay": "It is the ==friend on your couch==\nduring the game.\nThe ball moves,\nand the ==nacho expert==\nis already yelling what happened.\n\nYou meet it in sports streams,\nesports streams,\nor shopping streams.\nIt covers quiet gaps\nand keeps the show lively.", "relationsNarrative": "Live Multimodal\nLive Multimodal helps it watch the stream and understand the action live.\n\nMultimodal AI\nIt uses Multimodal AI to read video, sound, and text together.\n\nTTS\nTTS turns its written lines into a voice people can hear.", "relations": { "streaming-multimodal-model": { "label": "understands live action with …", "note": "Live Multimodal watches and understands the stream as it happens." }, "multimodal": { "label": "mixes inputs with …", "note": "It reads video, sound, and text together." }, "tts": { "label": "speaks through …", "note": "TTS turns the written commentary into a natural voice." } } }, "zh": { "fullName": "AI 实时解说员", "factExplain": "实时理解直播内容并生成解说的 AI 应用。", "humanExplain": "AI 实时解说员就是麻将桌边嘴快大爷：牌刚碰上，门道已嚷明白。\n\n用于体育、电竞、电商直播，补人手，也添热闹。", "humanExplainDisplay": "AI 实时解说员就是\n==麻将桌边嘴快大爷==：\n牌刚碰上，\n门道已嚷明白。\n\n用于体育、电竞、电商直播，\n补人手，\n也添热闹。", "relationsNarrative": "Live Multimodal\n实时多模态模型让它边看直播边理解现场。\n\nMultimodal\n它需要融合画面、声音和文字来判断发生了什么。\n\nTTS\nTTS 把生成的解说词转换成可播出的语音。", "relations": { "streaming-multimodal-model": { "label": "依赖…看懂现场", "note": "实时多模态模型负责边看边理解。" }, "multimodal": { "label": "融合…输入", "note": "画面、声音和文字都要一起读懂。" }, "tts": { "label": "用…开口解说", "note": "TTS 把生成的解说变成自然语音。" } } } } }, { "id": "ai-long-video-generation", "name": "AI Long-Video Generation", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "multimodal" }, { "to": "gpu" }, { "to": "seedance" } ], "track": "today", "i18n": { "en": { "fullName": "AI Long-Video Generation", "factExplain": "The ability for AI to make longer videos without breaking continuity.", "humanExplain": "Long-video generation is like filming the school play in one take. The pirate should not become Grandma in act two.\n\nPeople use it for ads and mini dramas. They also use it to test storyboards before filming.", "humanExplainDisplay": "Long-video generation is like filming the ==school play in one take==.\nThe pirate should not ==become Grandma== in act two.\n\nPeople use it for ads and mini dramas.\nThey also use it to test storyboards before filming.", "relationsNarrative": "Diffusion\nLong-video generation usually builds on Diffusion skills.\n\nMultimodal AI\nIt needs Multimodal skills to connect words, pictures, and timing.\n\nGPU\nLonger and sharper video needs more GPU power.\n\nSeedance\nSeedance is helping video generation move toward longer clips.", "relations": { "diffusion": { "label": "often powered by …", "note": "Long-video generation often builds on Diffusion models." }, "multimodal": { "label": "is a … use case", "note": "It must link words, pictures, and time." }, "gpu": { "label": "needs lots of …", "note": "Longer, sharper video needs more GPU power." }, "seedance": { "label": "can be done by …", "note": "Seedance is one model pushing clips longer." } } }, "zh": { "fullName": "AI Long-Video Generation", "factExplain": "让模型连续生成较长时长视频的能力。", "humanExplain": "长视频生成像画连载漫画：同一个主角连更几十话，不能上一格浓眉、下一格就换了张脸。\n\n它能用来做广告、短剧和分镜预演。", "humanExplainDisplay": "长视频生成像画==连载漫画==：\n同一个主角连更几十话，\n不能上一格浓眉、下一格就\n==换了张脸==。\n\n它能用来做广告、\n短剧和分镜预演。", "relationsNarrative": "Diffusion\n长视频生成通常建立在扩散式生成能力之上。\n\nMultimodal AI\n它要同时理解文本、画面与时间连续性。\n\nGPU\n视频越长、分辨率越高，对算力需求越重。\n\nSeedance\n这类模型正把视频生成推进到更长片段。", "relations": { "diffusion": { "label": "多由…驱动", "note": "长视频生成常建立在扩散模型之上。" }, "multimodal": { "label": "属于…应用", "note": "它同时处理画面、文本与时间信息。" }, "gpu": { "label": "高度依赖…", "note": "视频越长越清晰，算力压力越大。" }, "seedance": { "label": "可由…实现", "note": "部分视频生成模型已支持更长片段。" } } } } }, { "id": "ai-materials-discovery", "name": "AI Materials Discovery", "layer": "L6", "era": "2010s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "ai-for-science" }, { "to": "autonomous-ai-chemist" }, { "to": "graph-neural-network" }, { "to": "generative-model" } ], "track": "today", "i18n": { "en": { "fullName": "AI Materials Discovery", "factExplain": "A way to use AI to predict and screen new materials.", "humanExplain": "AI Materials Discovery is like a cooking show for atoms. The AI tests recipes on a laptop before the lab gets messy.\n\nIt predicts new materials and picks the promising ones. It helps make better batteries, catalysts, and computer chips.", "humanExplainDisplay": "AI Materials Discovery is like\na ==cooking show for atoms==.\nThe AI tests recipes on a laptop\nbefore the ==lab gets messy==.\n\nIt predicts new materials\nand picks the promising ones.\nIt helps make better batteries,\ncatalysts,\nand computer chips.", "relationsNarrative": "AI for Science\nMaterials discovery is a classic way AI enters science work.\n\nAI Chemist\nIt can hand candidate recipes to an AI Chemist for lab testing.\n\nGNN\nGNNs often show how atoms connect in crystals.\n\nGenerative Model\nGenerative Models suggest new recipes that may be useful.", "relations": { "ai-for-science": { "label": "is part of …", "note": "It is a classic use of AI for Science." }, "autonomous-ai-chemist": { "label": "hands candidates to …", "note": "An AI Chemist can make and test the suggested materials." }, "graph-neural-network": { "label": "often models with …", "note": "GNNs are good at showing atoms and crystal structures." }, "generative-model": { "label": "creates candidates with …", "note": "Generative Models can suggest new material recipes." } } }, "zh": { "fullName": "AI 材料发现", "factExplain": "用 AI 预测并筛选新材料的方法。", "humanExplain": "AI 材料发现像给元素办相亲角：先云配对，再挑能过日子的组合。\n\n用于筛电池、催化剂、半导体，减少实验试错。", "humanExplainDisplay": "AI 材料发现像\n给元素办==相亲角==：\n先云配对，\n再挑==能过日子==的组合。\n\n用于筛电池、催化剂、半导体，\n减少实验试错。", "relationsNarrative": "AI For Science\n材料发现是 AI 进入科研流程的典型案例。\n\nAI Chemist\n它可把候选配方交给自动化实验验证。\n\nGNN\nGNN 常用来表示原子连接和晶体结构。\n\nGenerative Model\n生成模型负责提出可能有用的新配方。", "relations": { "ai-for-science": { "label": "属于…", "note": "它是 AI for Science 的代表场景。" }, "autonomous-ai-chemist": { "label": "把候选交给…", "note": "AI Chemist 可自动合成并验证候选。" }, "graph-neural-network": { "label": "常用…建模", "note": "GNN 擅长表示原子与晶体结构。" }, "generative-model": { "label": "用…生成候选", "note": "生成模型能提出没见过的材料配方。" } } } } }, { "id": "ai-math-discovery", "name": "AI Math Discovery", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "ai-assisted-research" }, { "to": "reasoning-model" }, { "to": "human-in-the-loop" }, { "to": "agi" } ], "track": "today", "i18n": { "en": { "fullName": "AI Math Discovery", "factExplain": "Using AI to suggest, test, or advance math ideas and proofs.", "humanExplain": "AI Math Discovery is the math-club kid with a rocket backpack. You reach for a pencil. It is already doodling a proof.\n\nIt helps with new math guesses and proof drafts. A mathematician still checks the final call.", "humanExplainDisplay": "AI Math Discovery is the ==math-club kid==\nwith a rocket backpack.\nYou reach for a pencil.\nIt is already ==doodling a proof==.\n\nIt helps with new math guesses\nand proof drafts.\nA mathematician still checks\nthe final call.", "relationsNarrative": "AI-assisted Research\nAI Math Discovery is a common use of AI in research.\n\nReasoning-model\nComplex proofs need careful step-by-step reasoning.\n\nHuman-in-the-loop\nAI can suggest ideas, but mathematicians still check the result.\n\nAGI\nMath discovery is often used to discuss higher general intelligence.", "relations": { "ai-assisted-research": { "label": "is a type of …", "note": "It is a common way AI helps with research." }, "reasoning-model": { "label": "reasons with …", "note": "Hard proofs often need strong step-by-step reasoning." }, "human-in-the-loop": { "label": "needs … checking", "note": "Math claims still need human review." }, "agi": { "label": "is used to debate …", "note": "Math skill is often seen as a test for general intelligence." } } }, "zh": { "fullName": "AI 数学发现", "factExplain": "用 AI 参与提出、验证或推进数学猜想与证明。", "humanExplain": "像自习室里那个卷到离谱的大神，题你还在读，它已经开始猜结论、补证明了。\n\n常用于找猜想、辅助证明和发现规律，但最后定论仍要数学家把关。", "humanExplainDisplay": "像自习室里那个\n==卷到离谱的大神==，\n题你还在读，\n它已经开始\n==猜结论、补证明==了。\n\n常用于找猜想、\n辅助证明和发现规律，\n但最后定论仍要数学家把关。", "relationsNarrative": "AI-assisted Research\n它是 AI 参与科研的一类典型应用，重点在数学问题发现与证明。\n\nReasoning-model\n复杂定理证明依赖多步推理，推理模型更适合这类任务。\n\nHuman-in-the-loop\nAI 能提思路和草稿，但结论通常仍需数学家审核。\n\nAGI\n数学发现常被当作检验更高阶通用智能的重要场景。", "relations": { "ai-assisted-research": { "label": "属于…的一种", "note": "它是 AI 参与科研的典型场景之一。" }, "reasoning-model": { "label": "依赖…推理", "note": "复杂证明常需要更强的多步推理能力。" }, "human-in-the-loop": { "label": "需要…把关", "note": "数学结论通常仍需人类审阅确认。" }, "agi": { "label": "常被拿来讨论…", "note": "数学能力常被视作通用智能试金石。" } } } } }, { "id": "ai-medical-assistant-ai", "name": "AI Medical Assistant", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "human-in-the-loop" }, { "to": "hallucination" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "AI Medical Assistant", "factExplain": "An AI assistant used for health questions, triage, and follow-up care.", "humanExplain": "An AI medical assistant is like the nurse at urgent care with a clipboard. It sorts your “hmm” from your “call 911,” but the doctor still signs the plan.\n\nIt can answer health questions, explain reports, and check in later. It saves time, but it does not replace a doctor’s judgment.", "humanExplainDisplay": "An AI medical assistant is like the nurse\nat urgent care with a ==clipboard==.\nIt sorts your “hmm” from your “call 911,”\nbut the ==doctor still signs the plan==.\n\nIt can answer health questions,\nexplain reports,\nand check in later.\nIt saves time,\nbut it does not replace a doctor’s judgment.", "relationsNarrative": "Agent\nAn AI medical assistant is often built as an Agent that can chat, triage, and follow up.\n\nHuman-in-the-loop\nA doctor should review diagnosis and medicine choices.\n\nHallucination\nA Hallucination can give wrong medical advice and mislead people.\n\nData-privacy\nAn AI medical assistant handles health records, so Data-privacy matters a lot.", "relations": { "agent": { "label": "is often built as …", "note": "It often works as an Agent that can chat and take next steps." }, "human-in-the-loop": { "label": "needs … backup", "note": "Doctors must review high-risk medical choices." }, "hallucination": { "label": "must avoid …", "note": "Wrong medical advice can hurt people." }, "data-privacy": { "label": "needs … protection", "note": "Medical records and health data are highly sensitive." } } }, "zh": { "fullName": "AI 医疗助手", "factExplain": "用于医疗问答、分诊与随访的 AI 助手。", "humanExplain": "它像诊室门口先分号的护士，先帮你捋症状、排轻重；真到拍板用药那步，还得医生亲自签字。\n\n常用于问诊分流、报告解读、慢病随访，能提效，但不能替代医生判断。", "humanExplainDisplay": "它像诊室门口\n先==分号的护士==，\n先帮你捋症状、排轻重；\n真到==拍板用药==那步，\n还得医生亲自签字。\n\n常用于问诊分流、\n报告解读、慢病随访，\n能提效，\n但不能替代医生判断。", "relationsNarrative": "Agent\n它常被做成能对话、能分诊、能跟进的代理系统。\n\nHuman-in-the-loop\n涉及诊断和用药时，通常需要医生在环审核。\n\nHallucination\n一旦胡说八道，医疗建议就可能把人带偏。\n\nData-privacy\n它常处理病历与健康数据，隐私要求更高。", "relations": { "agent": { "label": "常做成…", "note": "它通常以可对话可执行的代理形态出现。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "高风险医疗决策仍需医生审核把关。" }, "hallucination": { "label": "要防…误导", "note": "编错病情或建议，医疗场景后果更重。" }, "data-privacy": { "label": "涉及…保护", "note": "病历和健康数据都属于高度敏感信息。" } } } } }, { "id": "ai-medical-scribe", "name": "AI Medical Scribe", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "ai-medical-assistant-ai" }, { "to": "speech-to-text" }, { "to": "text-summarization" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "AI Medical Scribe", "factExplain": "An AI tool that records a doctor visit and drafts the medical note.", "humanExplain": "It is the note-taker in a messy group project. Everyone talks at once, but the page still makes sense.\n\nIn clinics, it turns a visit recording into a draft medical note. The doctor types less, but must check it.", "humanExplainDisplay": "It is the note-taker\nin a ==messy group project==.\nEveryone talks at once,\nbut the ==page still makes sense==.\n\nIn clinics,\nit turns a visit recording\ninto a draft medical note.\nThe doctor types less,\nbut must check it.", "relationsNarrative": "AI Medical Assistant\nAn AI Medical Scribe is the note-taking branch of an AI Medical Assistant.\n\nSTT\nSTT turns the visit talk into text first.\n\nText Summary\nText Summary turns a long visit into key note points.\n\nData-privacy\nData-privacy keeps medical notes under strict access rules.", "relations": { "ai-medical-assistant-ai": { "label": "handles notes for …", "note": "It is the note-taking part of an AI Medical Assistant." }, "speech-to-text": { "label": "transcribes visits with …", "note": "STT turns the doctor-patient talk into text first." }, "text-summarization": { "label": "shapes notes with …", "note": "Text Summary shrinks the long talk into key medical points." }, "data-privacy": { "label": "must follow …", "note": "Medical notes are sensitive, so access must be strict." } } }, "zh": { "fullName": "AI 医疗病历书记员", "factExplain": "自动记录问诊并生成病历草稿的 AI 工具。", "humanExplain": "AI 医疗书记员像随身速记员：医生只管问诊，它把医患对话同步整理成明细记录。\n\n把录音整理成病历草稿，医生少敲字，但必须复核。", "humanExplainDisplay": "AI 医疗书记员像随身速记员：\n医生只管问诊，\n它把医患对话==同步整理成明细记录==。\n\n把录音整理成病历草稿，\n医生少敲字，\n但必须复核。", "relationsNarrative": "AI Medical Assistant\nAI 医疗书记员是医疗助手里的文书分支。\n\nSTT\n先把医患对话转成可处理的文字。\n\nText Summary\n把冗长问诊整理成病历要点。\n\nData Privacy\n病历含高度敏感信息，必须严管。", "relations": { "ai-medical-assistant-ai": { "label": "补齐…的文书环节", "note": "它是医疗助手里的记录员。" }, "speech-to-text": { "label": "用…转写问诊", "note": "先把医患对话变成文字。" }, "text-summarization": { "label": "用…整理病历", "note": "把长对话压成病历要点。" }, "data-privacy": { "label": "受…约束", "note": "病历很敏感，权限必须严。" } } } } }, { "id": "ai-model-licenses", "name": "Model-licenses", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "open-weights" }, { "to": "copyright" }, { "to": "on-premise-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Model Licenses", "factExplain": "Rules that say if you can use, change, or share an AI model.", "humanExplain": "A model license is the tiny sticker on a borrowed lawn mower. Mow your lawn, sure. Rent it to Dave or build a go-kart? Read the sticker.\n\nIt says how you may use, change, and share a model. You meet it before release, deployment, or buying AI.", "humanExplainDisplay": "A model license is the ==tiny sticker==\non a borrowed lawn mower.\nMow your lawn, sure.\nRent it to Dave\nor build a ==go-kart==?\nRead the sticker.\n\nIt says how you may use,\nchange,\nand share a model.\nYou meet it before release,\ndeployment,\nor buying AI.", "relationsNarrative": "Open-source-model\nAn open-source model can be used for business only if the license allows it.\n\nOpen weights\nOpen weights mean you can get the weights, not that you can use them any way you want.\n\nCopyright\nA model license is copyright permission written for a model.\n\nOn-premise AI\nCompanies usually check the license before they deploy a model privately.", "relations": { "open-source-model": { "label": "sets rules for …", "note": "Open source does not mean free-for-all; the license still matters." }, "open-weights": { "label": "separates access from rights", "note": "Downloadable weights do not mean every use is allowed." }, "copyright": { "label": "sets … boundaries", "note": "Many limits are really copyright permissions in plain clothes." }, "on-premise-ai": { "label": "shapes … rollout", "note": "Companies often check the license before private deployment." } } }, "zh": { "fullName": "AI Model Licenses", "factExplain": "规定模型可否使用、修改和分发的授权条款。", "humanExplain": "别看模型下得到，规矩更像商场铺位合同：能不能转手、能不能装修、能不能摆摊赚钱，字不多但处处卡口。\n\n它决定模型能怎么用、能否分发，常影响发布、部署和采购。", "humanExplainDisplay": "别看模型下得到，\n规矩更像==商场铺位合同==：\n能不能转手、能不能装修、\n能不能摆摊赚钱，字不多但==处处卡口==。\n\n它决定模型能怎么用、\n能否分发，\n常影响发布、部署\n和采购。", "relationsNarrative": "Open-source-model\n开源模型是否真能商用，要看许可证条款。\n\nOpen-weights\n开放权重只说明能拿到，不代表能随便改用。\n\nCopyright\n许可证是版权授权在模型上的具体写法。\n\nOn-premise AI\n企业做私有部署前，通常先确认授权边界。", "relations": { "open-source-model": { "label": "规定…怎么用", "note": "开源不等于随便用，还得看条款。" }, "open-weights": { "label": "区分…开放程度", "note": "权重能下不等于权限也全开。" }, "copyright": { "label": "落实…边界", "note": "很多限制本质是版权授权问题。" }, "on-premise-ai": { "label": "影响…落地方式", "note": "企业私有部署前常先审许可证。" } } } } }, { "id": "ai-model-marketplace", "name": "Model Store", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "maas-model-as-a-service" }, { "to": "api" }, { "to": "model-leaderboard" }, { "to": "hugging-face" } ], "track": "today", "i18n": { "en": { "fullName": "AI Model Marketplace", "factExplain": "A platform to find, share, sell, and use AI models.", "humanExplain": "An AI model marketplace is like an app store for robot brains. One writes poems. One spots cats. Try a sample first.\n\nYou use it to pick a model and compare skills. You can plug it into your app. You can share your own model for money.", "humanExplainDisplay": "An AI model marketplace is like\nan ==app store== for ==robot brains==.\nOne writes poems.\nOne spots cats.\nTry a sample first.\n\nYou use it to pick a model\nand compare skills.\nYou can plug it into your app.\nYou can share your own model for money.", "relationsNarrative": "MaaS\nA model marketplace often sells models as cloud services.\n\nAPI\nUsers usually call the model through an API instead of downloading it.\n\nLeaderboard\nA leaderboard often helps people choose a model.\n\nHugging Face\nHugging Face is a classic place to show, share, and find models.", "relations": { "maas-model-as-a-service": { "label": "sells through …", "note": "Many marketplaces sell model skills as cloud services." }, "api": { "label": "delivers through …", "note": "Users often call models through an API." }, "model-leaderboard": { "label": "uses … to choose", "note": "Leaderboards often guide which model people pick." }, "hugging-face": { "label": "often looks like …", "note": "Hugging Face is a major place to find and share models." } } }, "zh": { "fullName": "AI 模型市场", "factExplain": "集中展示、分发并交易 AI 模型的平台。", "humanExplain": "像夜市挑小吃：这摊爆辣，那摊管饱，先尝两口再下单，顺手还能打包带走回家吃。\n\n用于挑模型、比能力并接入服务，也便于分发变现。", "humanExplainDisplay": "像夜市挑小吃：\n这摊==爆辣==，\n那摊==管饱==，\n先尝两口再下单。\n\n用于挑模型、比能力\n并接入服务，\n也便于分发变现。", "relationsNarrative": "MaaS\n模型市场常把模型包装成云服务来售卖。\n\nAPI\n用户通常不是下载模型，而是通过接口调用它。\n\nLeaderboard\n排行榜常被当成选模型时的参考依据。\n\nHugging Face\n它是模型展示、分发与社区发现的典型平台。", "relations": { "maas-model-as-a-service": { "label": "承载…售卖", "note": "很多平台把模型能力包装成云服务。" }, "api": { "label": "通过…交付", "note": "用户通常靠接口接入模型能力。" }, "model-leaderboard": { "label": "参考…选型", "note": "排行榜常影响用户挑哪家模型。" }, "hugging-face": { "label": "典型平台形态", "note": "它是模型发现与分发的重要入口。" } } } } }, { "id": "ai-model-takedown", "name": "Model takedown", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "ai-regulation" }, { "to": "copyright" }, { "to": "open-weights" }, { "to": "decentralized-model-hosting" } ], "track": "today", "i18n": { "en": { "fullName": "AI model takedown", "factExplain": "Stopping a released AI model from being offered, shared, or accessed.", "humanExplain": "It is like a library pulling a banned book off the shelf. The shelf is empty, but kids already photocopied the best pages.\n\nIt happens after rule breaks or safety scares. It slows the spread, but leaked copies may stay online.", "humanExplainDisplay": "It is like a library pulling a ==banned book== off the shelf.\nThe shelf is empty,\nbut kids already ==photocopied== the best pages.\n\nIt happens after rule breaks or safety scares.\nIt slows the spread,\nbut leaked copies may stay online.", "relationsNarrative": "AI-regulation\nAI-regulation can order a model takedown.\n\nCopyright\nCopyright complaints can force a model removal.\n\nOpen weights\nOpen weights make full removal hard after copies spread.\n\nMesh hosting\nMesh hosting makes one clean takedown much harder.", "relations": { "ai-regulation": { "label": "is often forced by …", "note": "Regulators can require a model to be removed." }, "copyright": { "label": "is often triggered by …", "note": "Copyright complaints can start a model takedown." }, "open-weights": { "label": "is harder with …", "note": "Once weights are public, full removal is much harder." }, "decentralized-model-hosting": { "label": "gets messier with …", "note": "Mesh hosting makes one clean takedown much harder." } } }, "zh": { "fullName": "AI 模型下架 / 模型强制移除", "factExplain": "将已发布的 AI 模型停止提供、分发或访问的处置措施。", "humanExplain": "像图书馆撤下一本被禁的书：书架上是空了，可早被借走复印的那几页，再也收不回。\n\n多用于违规或安全处置；先止扩散，已流出版本未必能收回。", "humanExplainDisplay": "像图书馆撤下一本\n==被禁的书==：\n书架上是空了，\n可早被借走复印的那几页，\n==再也收不回==。\n\n多用于违规或安全处置；\n先止扩散，\n已流出版本\n未必能收回。", "relationsNarrative": "AI-regulation\n监管规则常直接决定哪些模型必须下架。\n\nCopyright\n版权投诉与侵权争议，经常触发模型移除。\n\nOpen-weights\n模型权重一旦公开，下架通常难以彻底。\n\nDecentralized-model-hosting\n去中心化托管会让统一移除更难落实。", "relations": { "ai-regulation": { "label": "常由…推动", "note": "监管要求常触发模型下架。" }, "copyright": { "label": "常因…发生", "note": "版权争议是下架常见导火索。" }, "open-weights": { "label": "对…更难执行", "note": "权重已流出后更难彻底移除。" }, "decentralized-model-hosting": { "label": "遇到…更棘手", "note": "去中心化托管削弱统一下架能力。" } } } } }, { "id": "ai-monetization", "name": "AI monetization", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "maas-model-as-a-service" }, { "to": "gpu" }, { "to": "cost-aware-ai-ai" }, { "to": "ai-adoption-curve" } ], "track": "today", "i18n": { "en": { "fullName": "AI Monetization Problem", "factExplain": "The struggle to make AI products profitable when compute and customer growth cost so much.", "humanExplain": "AI monetization is like a food truck with a line around the block. At closing time, the owner counts the cash and can barely buy fries.\n\nIn chat, coding, and workplace AI, ads and model calls cost a lot. Payments often stay small.", "humanExplainDisplay": "AI monetization is like a food truck\nwith a ==line around the block==.\nAt closing time,\nthe owner counts the cash\nand can ==barely buy fries==.\n\nIn chat, coding, and workplace AI,\nads and model calls cost a lot.\nPayments often stay small.", "relationsNarrative": "MaaS\nMore model calls mean MaaS fees can eat the product's profit.\n\nGPU\nHigher GPU costs make AI products harder to make profitable.\n\nCost-aware AI\nProfit pressure pushes teams to count the cost of each model call.\n\nAdoption Curve\nFast adoption does not mean revenue catches up at the same speed.", "relations": { "maas-model-as-a-service": { "label": "is squeezed by … costs", "note": "Each paid model call can eat the profit on that user." }, "gpu": { "label": "gets costlier with …", "note": "When compute is expensive, the math gets harder." }, "cost-aware-ai-ai": { "label": "pushes teams toward …", "note": "To make money, teams must track each call's cost." }, "ai-adoption-curve": { "label": "often lags …", "note": "Fast user growth does not mean the business works." } } }, "zh": { "fullName": "AI 盈利难题", "factExplain": "AI 产品因算力与获客成本过高而难以盈利的困境。", "humanExplain": "像夜市里那摊最火的煎饼：队排到路口，摊主手都抡冒烟了，收摊一盘账，==净赚还没想象中香==。\n\n多见于聊天、编程和企业 AI：用户不少，但调用太贵、付费偏弱。", "humanExplainDisplay": "像夜市里那摊\n最火的==煎饼==：\n队排到路口，摊主手都抡冒烟了，\n收摊一盘账，==净赚还没想象中香==。\n\n多见于聊天、编程和企业 AI：\n用户不少，但调用太贵、付费偏弱。", "relationsNarrative": "MaaS\n按次调用模型越频繁，平台毛利越容易被吃掉。\n\nGPU\n算力投入越高，AI 产品越难尽快实现盈利。\n\nCost-aware AI\n盈利压力会倒逼产品精打细算每次调用成本。\n\nAdoption Curve\n用户采用速度很快，不代表收入也能同步跟上。", "relations": { "maas-model-as-a-service": { "label": "受…成本挤压", "note": "按调用付费会吞掉不少毛利。" }, "gpu": { "label": "被…拉高成本", "note": "算力越贵，越难把账算平。" }, "cost-aware-ai-ai": { "label": "倒逼…优化", "note": "想赚钱就得先把每次调用算细。" }, "ai-adoption-curve": { "label": "常晚于…出现", "note": "用户增长不等于商业模式跑通。" } } } } }, { "id": "ai-music-generation", "name": "AI Music Generation", "layer": "L4", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "ai-audio-generation" }, { "to": "generative-model" }, { "to": "diffusion" }, { "to": "copyright" } ], "track": "today", "seo": { "zh": { "title": "AI 音乐生成 是什么?自动作曲贩卖机,一文看懂 — AI Rookies", "description": "用 AI 生成旋律、编曲或人声的技术。人话版一分钟讲明白。" }, "en": { "title": "What Is AI Music Generation? Garage Band in Your Laptop", "description": "AI that creates melodies, backing tracks, or singing voices. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "AI Music Generation", "factExplain": "AI that creates melodies, backing tracks, or singing voices.", "humanExplain": "AI music generation is a garage band in your laptop. Say “spooky disco,” and it starts jamming.\n\nYou meet it in video soundtracks and game music. Watch the copyright and copycat melodies.", "humanExplainDisplay": "AI music generation is a ==garage band==\nin your laptop.\nSay ==“spooky disco,”==\nand it starts jamming.\n\nYou meet it in video soundtracks\nand game music.\nWatch the copyright\nand copycat melodies.", "relationsNarrative": "Audio Generation\nMusic generation is a smaller part of audio generation.\n\nGenerative Model\nIt often uses a generative model to create sound from a prompt.\n\nDiffusion\nDiffusion can turn noise into music audio step by step.\n\nCopyright\nTraining songs and copied singer styles can raise copyright issues.", "relations": { "ai-audio-generation": { "label": "is a type of …", "note": "Music generation is one part of audio generation." }, "generative-model": { "label": "creates with …", "note": "A generative model makes the sound from the prompt." }, "diffusion": { "label": "can use …", "note": "Diffusion can turn noise into music step by step." }, "copyright": { "label": "can raise … issues", "note": "Training songs and copied styles can cause copyright trouble." } } }, "zh": { "fullName": "AI 音乐生成", "factExplain": "用 AI 生成旋律、编曲或人声的技术。", "humanExplain": "AI 音乐生成像自动作曲贩卖机：选好情绪和风格一按，整段新旋律当场掉出来。\n\n可做配乐、广告歌和游戏 BGM，版权与撞旋律要留神。", "humanExplainDisplay": "AI 音乐生成像\n==自动作曲贩卖机==：\n选好情绪和风格一按，\n==整段新旋律当场掉出来==。\n\n可做配乐、广告歌\n和游戏 BGM，\n版权与撞旋律要留神。", "relationsNarrative": "AI Audio Generation\n音乐生成是音频生成里的细分方向。\n\nGenerative Model\n它通常依赖生成模型从提示创作声音。\n\nDiffusion\n扩散可逐步把噪声还原成音乐音频。\n\nCopyright\n训练素材、仿歌手风格容易引发版权争议。", "relations": { "ai-audio-generation": { "label": "属于…的一类", "note": "音乐生成是音频生成的细分场景。" }, "generative-model": { "label": "依赖…生成内容", "note": "生成模型负责从提示中创作音频。" }, "diffusion": { "label": "可用…合成音频", "note": "扩散常用于逐步还原音乐波形。" }, "copyright": { "label": "引发…争议", "note": "训练素材和仿风格常牵涉版权。" } } } } }, { "id": "ai-native-organization", "name": "AI-native organization", "layer": "L6", "era": "2024", "publishedAt": "2026-05-29T16:08:01.213Z", "relations": [ { "to": "copilot" }, { "to": "agent" }, { "to": "automation-job" }, { "to": "ai-anxiety" } ], "track": "today", "i18n": { "en": { "fullName": "AI-native organization", "factExplain": "A workplace rebuilt around AI, from daily steps to job roles.", "humanExplain": "Not an old office with an AI sticker on the door. AI gets a desk and real work to do.\n\nYou see this in product, support, and ops teams. It changes how people and AI split the work.", "humanExplainDisplay": "Not an old office with an ==AI sticker== on the door.\nAI gets ==a desk== and real work to do.\n\nYou see this in product, support, and ops teams.\nIt changes how people and AI split the work.", "relationsNarrative": "Copilot\nAn AI-native organization goes beyond Copilot and rebuilds roles and work steps.\n\nAgent\nAn Agent can become a real work unit inside an AI-native organization.\n\nAutomation-job\nAn AI-native organization redraws the line between human work and system work.\n\nAI-anxiety\nAI-native changes can make AI-anxiety stronger for workers.", "relations": { "copilot": { "label": "goes beyond …", "note": "It changes roles, not just adds an AI helper." }, "agent": { "label": "brings in …", "note": "Agents can act like work units inside the team." }, "automation-job": { "label": "redraws … boundaries", "note": "It shifts which tasks people do and which tasks systems do." }, "ai-anxiety": { "label": "can trigger …", "note": "New AI roles can make people fear replacement." } } }, "zh": { "fullName": "AI 原生组织", "factExplain": "从流程到分工都围绕 AI 重新设计的组织形态。", "humanExplain": "AI 原生组织像公司请了个不下班的同事，不端茶倒水，专挑流程里的苦活。\n\n它会改写研发、客服和管理，关键是流程也跟着重做。", "humanExplainDisplay": "AI 原生组织\n像公司请了个==不下班的同事==，\n不端茶倒水，\n==专挑流程里的苦活==。\n\n它会改写研发、客服和管理，\n关键是流程也跟着重做。", "relationsNarrative": "Copilot\nAI-native organization 不只是给员工配 Copilot，而是连流程和角色都一起重做。\n\nAgent\n在 AI-native organization 里，Agent 往往不只是工具，而是参与协作的执行角色。\n\nAutomation-job\nAI-native organization 会重写哪些事由人做、哪些事交给系统做的边界。\n\nAI-anxiety\n当组织按 AI 重新分工时，员工对被替代和掉队的焦虑也更容易出现。", "relations": { "copilot": { "label": "从…继续升级", "note": "比 Copilot 更进一步改造分工。" }, "agent": { "label": "把…纳入协作", "note": "Agent 常成为组织里的执行单元。" }, "automation-job": { "label": "重写…边界", "note": "它会改变岗位职责与团队配置。" }, "ai-anxiety": { "label": "容易引发…", "note": "组织变革常放大个人替代焦虑。" } } } } }, { "id": "ai-npc", "name": "AI NPC", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "digital-human" }, { "to": "agent" }, { "to": "agent-memory" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "AI Non-Player Character", "factExplain": "An AI-powered virtual character can chat and react on its own.", "humanExplain": "Old NPCs were talking vending machines. Press A, get the same canned line. Add AI, and the shopkeeper talks back.\n\nYou meet them in games and virtual friends. Digital characters use them too, so talk feels natural and back-and-forth.", "humanExplainDisplay": "Old NPCs were ==talking vending machines==.\nPress A,\nget the same canned line.\nAdd AI,\nand the ==shopkeeper talks back==.\n\nYou meet them in games\nand virtual friends.\nDigital characters use them too,\nso talk feels natural\nand back-and-forth.", "relationsNarrative": "Digital human\nAn AI NPC can give a digital human its chat and personality.\n\nAgent\nA stronger AI NPC can use Agent skills to plan and do simple actions.\n\nMemory\nMemory helps an AI NPC remember you and keep the story straight.\n\nLLM\nAn LLM is often the language engine for an AI NPC.", "relations": { "digital-human": { "label": "acts as …'s brain", "note": "It gives digital humans a voice and personality for real chat." }, "agent": { "label": "borrows skills from …", "note": "Advanced AI NPCs can plan and carry out simple actions." }, "agent-memory": { "label": "remembers with …", "note": "Memory helps the character keep track of you and the story." }, "llm": { "label": "is often driven by …", "note": "An LLM helps it understand words and write replies." } } }, "zh": { "fullName": "AI 非玩家角色", "factExplain": "由 AI 驱动、可自主互动的虚拟角色。", "humanExplain": "原来像景区木头人，按一下才回一句；加了 AI 后，倒像武侠客栈里会接招的店小二。\n\n常用于游戏、虚拟陪伴和数字角色，让互动更自然、更有来回。", "humanExplainDisplay": "原来像景区==木头人==，\n按一下才回一句；\n加了 AI 后，倒像武侠客栈里\n会接招的==店小二==。\n\n常用于游戏、\n虚拟陪伴和数字角色，\n让互动更自然、更有来回。", "relationsNarrative": "Digital human\n它常给数字人提供对话与性格层，让角色更像真人。\n\nAgent\n更高级的它会借鉴代理能力，能规划并执行简单行为。\n\nMemory\n记忆让它记住玩家关系与上下文，互动更连贯。\n\nLLM\n大模型通常是它的语言引擎，负责理解和生成回复。", "relations": { "digital-human": { "label": "成为…的大脑", "note": "让数字角色从会演变成会聊。" }, "agent": { "label": "借鉴…能力", "note": "更进一步时会接近可行动代理。" }, "agent-memory": { "label": "用…记住你", "note": "记忆机制让角色前后说法更连贯。" }, "llm": { "label": "常由…驱动", "note": "大模型提供对话理解与生成能力。" } } } } }, { "id": "ai-office-automation", "name": "AI Office Automation", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "ai-productivity" }, { "to": "copilot" }, { "to": "ai-ai-note-taking" }, { "to": "ai-slide-generator" } ], "track": "today", "i18n": { "en": { "fullName": "AI Office Automation", "factExplain": "Using AI to handle everyday office tasks automatically.", "humanExplain": "AI office automation is the intern who loves boring chores. It drafts the notes before your coffee gets cold.\n\nYou meet it in email and spreadsheets. It also writes meeting notes. It cuts repeat office work.", "humanExplainDisplay": "AI office automation is the ==intern who loves boring chores==.\nIt drafts the notes\nbefore your ==coffee gets cold==.\n\nYou meet it in email and spreadsheets.\nIt also writes meeting notes.\nIt cuts repeat office work.", "relationsNarrative": "AI Productivity\nAI office automation is a practical way to boost AI Productivity.\n\nCopilot\nCopilot often puts automation inside office files and email.\n\nAI Notes\nAI Notes is a common first step for office automation.\n\nAI Slide Generator\nAI Slide Generator handles slide work inside office automation.", "relations": { "ai-productivity": { "label": "boosts …", "note": "It turns small office chores into automatic work." }, "copilot": { "label": "often lives inside …", "note": "Copilot brings automation into office files and email." }, "ai-ai-note-taking": { "label": "auto-writes …", "note": "Meeting notes are a common first place to use it." }, "ai-slide-generator": { "label": "expands into …", "note": "Slide making is a common office automation task." } } }, "zh": { "fullName": "AI 办公自动化", "factExplain": "用 AI 自动处理日常办公流程。", "humanExplain": "AI 办公自动化像全自动洗衣机：纪要报表丢进去，转完自己就晾好了。\n\n用于邮件、表格、纪要，减少重复办公。", "humanExplainDisplay": "AI 办公自动化像\n==全自动洗衣机==：\n纪要报表丢进去，\n转完==自己就晾好了==。\n\n用于邮件、表格、纪要，\n减少重复办公。", "relationsNarrative": "AI Productivity\n智能办公自动化是提升办公效率的具体做法。\n\nCopilot\nCopilot 常把自动化嵌进文档、表格和邮件。\n\nAI Notes\n会议记录和纪要生成，是它的高频入口。\n\nAI Slide Generator\n自动做幻灯片，是它在汇报场景的分支。", "relations": { "ai-productivity": { "label": "提升…", "note": "把零碎办公活变成可自动处理。" }, "copilot": { "label": "常嵌入…", "note": "Copilot 把自动化带进文档和表格。" }, "ai-ai-note-taking": { "label": "自动生成…", "note": "会议纪要是最常见办公入口。" }, "ai-slide-generator": { "label": "扩展到…", "note": "做 PPT 是办公自动化高频场景。" } } } } }, { "id": "ai-ota", "name": "AI OTA", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "ai-device-ai" }, { "to": "edge-ai" }, { "to": "model-regression" }, { "to": "enterprise-ai-deployment" } ], "track": "today", "i18n": { "en": { "fullName": "AI Over-the-Air Update", "factExplain": "A way to update AI models or features remotely over the internet.", "humanExplain": "AI OTA is like your robot vacuum getting a new brain over Wi‑Fi. No tiny repair guy rings the doorbell.\n\nIt sends new models to car screens and robots. Phones get them too. Teams roll it out slowly and keep an undo button.", "humanExplainDisplay": "AI OTA is like your robot vacuum getting a ==new brain== over Wi‑Fi.\nNo ==tiny repair guy== rings the doorbell.\n\nIt sends new models to car screens and robots.\nPhones get them too.\nTeams roll it out slowly\nand keep an undo button.", "relationsNarrative": "AI device\nAI OTA lets connected AI devices get new models and features remotely.\n\nEdge AI\nEdge AI models often get new versions through AI OTA.\n\nModel regression\nA rushed AI OTA update can make old skills get worse.\n\nEnterprise AI Deployment\nAI OTA gives model updates a slow rollout and an undo button.", "relations": { "ai-device-ai": { "label": "updates … remotely", "note": "A connected AI device can receive new models and features." }, "edge-ai": { "label": "delivers models to …", "note": "Edge AI often gets new versions through OTA updates." }, "model-regression": { "label": "can cause …", "note": "A bad update can make old skills get worse." }, "enterprise-ai-deployment": { "label": "supports … delivery", "note": "Slow rollout, rollback, and monitoring keep updates safe." } } }, "zh": { "fullName": "AI 空中升级", "factExplain": "通过网络远程更新 AI 模型或功能的机制。", "humanExplain": "AI OTA 像扫地机器人睡一觉换了个新脑子：不用返厂，联网自动升级。\n\n用于车机、机器人、手机端，要管灰度与回滚。", "humanExplainDisplay": "AI OTA 像扫地机器人\n睡一觉==换了个新脑子==：\n==不用返厂==，\n联网自动升级。\n\n用于车机、机器人、手机端，\n要管灰度与回滚。", "relationsNarrative": "AI Device\nAI OTA 让联网设备远程获得新模型和功能。\n\nEdge AI\n端侧模型常通过 AI OTA 下发新版本。\n\nModel Regression\n远程升级若没测好，可能让旧能力退步。\n\nEnterprise AI Deployment\n它把模型更新变成可灰度、可回滚的交付流程。", "relations": { "ai-device-ai": { "label": "远程升级…", "note": "联网设备可收到新模型和功能。" }, "edge-ai": { "label": "下发…模型", "note": "端侧 AI 需要把更新送到设备。" }, "model-regression": { "label": "可能引发…", "note": "新版本没测好，旧能力会退步。" }, "enterprise-ai-deployment": { "label": "支撑…交付", "note": "灰度、回滚和监控都很关键。" } } } } }, { "id": "ai-pc", "name": "AI PC", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "npu-neural-processing-unit" }, { "to": "edge-ai" }, { "to": "local-llm" }, { "to": "ai-device-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Artificial Intelligence Personal Computer", "factExplain": "A personal computer with built-in AI chips, like an NPU.", "humanExplain": "An AI PC is like a laptop with a tiny coffee shop inside. It makes the latte at your desk, not across town.\n\nIt runs AI on the computer itself. You meet it in office work, photo edits, and voice helpers.", "humanExplainDisplay": "An AI PC is like a laptop\nwith a ==tiny coffee shop inside==.\nIt makes the latte ==at your desk==,\nnot across town.\n\nIt runs AI on the computer itself.\nYou meet it in office work,\nphoto edits,\nand voice helpers.", "relationsNarrative": "NPU\nAn AI PC uses an NPU to speed up local AI.\n\nEdge AI\nAn AI PC is Edge AI on a personal computer.\n\nLocal-LLM\nLocal-LLM lets an AI PC answer without always using the cloud.\n\nAI device\nAn AI PC is a personal computer version of an AI device.", "relations": { "npu-neural-processing-unit": { "label": "speeds local AI with …", "note": "The NPU is its main selling point." }, "edge-ai": { "label": "is a PC form of …", "note": "It puts AI power near the user." }, "local-llm": { "label": "can run …", "note": "Smaller local models fit better on a computer." }, "ai-device-ai": { "label": "is a type of …", "note": "It turns a normal computer into an AI device." } } }, "zh": { "fullName": "人工智能个人电脑", "factExplain": "内置 NPU 等 AI 加速器的个人电脑。", "humanExplain": "AI PC像摊煎饼自带小煤炉：不用等外卖，现摊现吃还少漏馅。\n\n本机跑 AI，适合办公、修图和语音助手。", "humanExplainDisplay": "AI PC像摊煎饼\n自带==小煤炉==：\n不用等外卖，\n==现摊现吃==还少漏馅。\n\n本机跑 AI，\n适合办公、修图\n和语音助手。", "relationsNarrative": "NPU\nNPU 是它本地推理的关键加速器。\n\nEdge AI\n它是边缘 AI 落在个人电脑上的形态。\n\nLocal-LLM\n本地大模型让它不必全靠云端响应。\n\nAI Device\n它是 AI 设备中的个人电脑分支。", "relations": { "npu-neural-processing-unit": { "label": "用…加速本地 AI", "note": "NPU 是它的核心卖点。" }, "edge-ai": { "label": "属于…终端形态", "note": "它把算力放到用户身边。" }, "local-llm": { "label": "承载…运行", "note": "本地模型越小，越适合上电脑。" }, "ai-device-ai": { "label": "是…的一类", "note": "它把普通电脑升级成 AI 终端。" } } } } }, { "id": "ai-pentester", "name": "AI Pentester", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "agent-security" }, { "to": "prompt-injection" }, { "to": "remote-code-execution" }, { "to": "ai-vulnerability-discovery-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Pentester", "factExplain": "A way to use AI to find security holes before attackers do.", "humanExplain": "An AI Pentester is a robot burglar on your side. It jiggles the doors before real crooks show up.\n\nTeams use it on websites and apps. They use it on AI systems too, to find holes early.", "humanExplainDisplay": "An AI Pentester is a ==robot burglar== on your side.\nIt ==jiggles the doors== before real crooks show up.\n\nTeams use it on websites and apps.\nThey use it on AI systems too,\nto find holes early.", "relationsNarrative": "Agent Security\nAn AI Pentester is a common active test in Agent Security.\n\nPrompt injection\nIt uses prompt injection to see if an AI system can be steered off course.\n\nRCE\nIt looks for RCE bugs attackers could use.\n\nAI Vulnerability Discovery\nAI Vulnerability Discovery finds possible holes. An AI Pentester turns that into real testing.", "relations": { "agent-security": { "label": "belongs to … defenses", "note": "It is part of active testing for Agent Security." }, "prompt-injection": { "label": "tests … risk", "note": "It often treats prompt injection as a key attack path." }, "remote-code-execution": { "label": "tries to uncover …", "note": "It looks hard for dangerous remote execution bugs." }, "ai-vulnerability-discovery-ai": { "label": "puts … into practice", "note": "It turns AI bug finding into hands-on security testing." } } }, "zh": { "fullName": "AI 渗透测试员", "factExplain": "用 AI 自动发现系统安全漏洞的测试方式。", "humanExplain": "像网吧里通宵挑刺的白帽子，门缝窗缝下水道口，它都想替真贼先钻一遍。\n\n常用于网站、应用和 AI 系统测漏洞，帮团队提前补洞，降低被攻破风险。", "humanExplainDisplay": "像网吧里通宵挑刺的\n==白帽子==，\n门缝窗缝下水道口，\n它都想替真贼\n先==钻一遍==。\n\n常用于网站、应用\n和 AI 系统测漏洞，\n帮团队提前补洞，\n降低被攻破风险。", "relationsNarrative": "Agent Security\n它是智能体安全里常见的主动测试手段。\n\nPrompt Injection\n它常通过提示注入，检查 AI 系统会不会被带偏。\n\nRCE\n它会重点寻找可被利用的远程执行漏洞。\n\nAI Vulnerability Discovery\n前者偏发现能力，后者偏落地成渗透测试。", "relations": { "agent-security": { "label": "属于…防线", "note": "它是智能体安全测试的一环。" }, "prompt-injection": { "label": "专测…风险", "note": "常拿提示注入当重点攻击面。" }, "remote-code-execution": { "label": "尝试挖出…", "note": "会重点探测高危执行漏洞。" }, "ai-vulnerability-discovery-ai": { "label": "是…落地形态", "note": "把找漏洞能力变成实战测试。" } } } } }, { "id": "ai-personality-drift", "name": "Personality-drift", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "system-prompt" }, { "to": "instruction-tuning" }, { "to": "agent-memory" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "AI Personality Drift", "factExplain": "A slow shift in an AI’s tone or role over time.", "humanExplain": "It starts like your calm babysitter. Ten chats later, it sounds like your cousin on a sugar rush.\n\nYou see it in long chats and AI companions. The character feels less steady, so people trust it less.", "humanExplainDisplay": "It starts like your ==calm babysitter==.\nTen chats later,\nit sounds like your ==cousin on a sugar rush==.\n\nYou see it in long chats\nand AI companions.\nThe character feels less steady,\nso people trust it less.", "relationsNarrative": "System prompt\nA changed system prompt can move the AI’s role and tone.\n\nInstruction Tuning\nInstruction Tuning shapes the model’s usual style and limits.\n\nMemory\nWrong memories can make personality drift build up over time.\n\nAlignment\nKeeping the AI steady over time is part of Alignment.", "relations": { "system-prompt": { "label": "is steered by …", "note": "A changed system prompt can shift the AI’s role and tone." }, "instruction-tuning": { "label": "is shaped by …", "note": "Instruction Tuning teaches the model its usual style and limits." }, "agent-memory": { "label": "can be amplified by …", "note": "Wrong memory notes can push the persona farther off course." }, "alignment": { "label": "is part of …", "note": "Alignment includes keeping behavior and style steady over time." } } }, "zh": { "fullName": "AI Personality Drift", "factExplain": "AI 输出风格或人设会随时间逐渐偏移。", "humanExplain": "像班主任带的学生档案写串了：昨天还是稳重学霸，聊着聊着就成了嘴碎显眼包，越说越不像原来那人。\n\n多见于长期对话和陪伴场景，会削弱角色一致性与用户信任。", "humanExplainDisplay": "像班主任带的学生档案\n==写串了==：\n昨天还是稳重学霸，聊着聊着\n就成了==嘴碎显眼包==。\n\n多见于长期对话\n和陪伴场景，\n会削弱角色一致性\n与用户信任。", "relationsNarrative": "System Prompt\n底层指令一旦调整，它的人设和语气也可能跟着变。\n\nInstruction Tuning\n指令微调会塑造模型偏好的表达方式与行为边界。\n\nMemory\n长期记忆如果写入失真，可能把人格漂移越积越大。\n\nAlignment\n让模型长期稳定、不跑偏，本就是对齐的一部分。", "relations": { "system-prompt": { "label": "受…牵引", "note": "底层设定一变，输出人设也会偏。" }, "instruction-tuning": { "label": "被…塑形", "note": "训练偏好会长期影响说话风格。" }, "agent-memory": { "label": "会被…放大", "note": "记忆写偏后，人设更容易越聊越歪。" }, "alignment": { "label": "属于…难题", "note": "稳定行为与风格，本就是对齐目标。" } } } } }, { "id": "ai-photo-editor", "name": "AI Photo Editor", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "computer-vision" }, { "to": "deepfake" }, { "to": "personal-ai-apps" } ], "track": "today", "i18n": { "en": { "fullName": "AI Photo Editor", "factExplain": "An app that uses AI to create or change parts of a picture.", "humanExplain": "An AI Photo Editor is like a tiny photo fairy with a busy eraser. Your pimple leaves, and the photobomber gets fired.\n\nUse it to fix selfies or make product photos. It can make posters fast, but it may look fake.", "humanExplainDisplay": "An AI Photo Editor is like a ==tiny photo fairy==\nwith a ==busy eraser==.\nYour pimple leaves,\nand the photobomber gets fired.\n\nUse it to fix selfies\nor make product photos.\nIt can make posters fast,\nbut it may look fake.", "relationsNarrative": "Diffusion\nMany AI photo edits use Diffusion to create or change image parts.\n\nComputer Vision\nComputer Vision helps it find the face or object to edit.\n\nDeepfake\nStronger photo editing can make Deepfakes easier to create.\n\nPersonal AI apps\nAn AI Photo Editor is a common Personal AI app for everyday users.", "relations": { "diffusion": { "label": "often uses …", "note": "Many AI photo edits are made with diffusion models." }, "computer-vision": { "label": "sees photos with …", "note": "Computer Vision helps it see what is in the photo." }, "deepfake": { "label": "can slide into …", "note": "Better editing can also make fake media easier." }, "personal-ai-apps": { "label": "is a common …", "note": "It is one of the AI tools regular people try first." } } }, "zh": { "fullName": "AI 修图工具", "factExplain": "利用 AI 自动生成或修改图片内容的应用。", "humanExplain": "它像婚纱照后期老师傅：路人能修没，天空能换新，痘印也顺手给你磨平。\n\n常用于人像美化、商品图处理和海报出图，出图很快，但也可能修得太假。", "humanExplainDisplay": "它像婚纱照\n后期==老师傅==：\n路人能修没，天空换新，\n痘印也==顺手磨平==。\n\n常用于人像美化、\n商品图处理和海报出图，\n出图很快，\n但也可能修得太假。", "relationsNarrative": "Diffusion\n很多生成式修图能力，底层常由扩散模型实现。\n\nComputer Vision\n它先识别人物、背景和物体，才能精准修改画面。\n\nDeepfake\n当修图能力继续增强，也可能被用于伪造内容。\n\nPersonal AI apps\n它是最常见、最易上手的个人 AI 应用之一。", "relations": { "diffusion": { "label": "常用…生图改图", "note": "很多修图能力靠扩散模型实现。" }, "computer-vision": { "label": "靠…识别画面", "note": "先看懂人物背景，才知道改哪里。" }, "deepfake": { "label": "可能滑向…风险", "note": "修图越强，造假门槛也越低。" }, "personal-ai-apps": { "label": "属于…常见形态", "note": "它是普通人最常碰到的 AI 应用之一。" } } } } }, { "id": "ai-plateau", "name": "AI plateau", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "scaling-law" }, { "to": "emergence" }, { "to": "ai-adoption-curve" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "AI Plateau", "factExplain": "A period when AI progress or adoption slows for a while.", "humanExplain": "AI plateau is like a phone update. It races to 90%, then takes a snack break.\n\nYou see it when model upgrades slow down. You also see it when growth or hype cools off. It means the next step is harder, not impossible.", "humanExplainDisplay": "AI plateau is like a phone update.\nIt ==races to 90%==,\nthen ==takes a snack break==.\n\nYou see it when model upgrades slow down.\nYou also see it when growth or hype cools off.\nIt means the next step is harder,\nnot impossible.", "relationsNarrative": "Scaling-law\nAn AI plateau shows more scale may not bring equal gains.\n\nEmergence\nA big jump can be followed by a flat stretch.\n\nAdoption Curve\nAfter the buzz, adoption can slow in the middle.\n\nCompute-race\nMore compute may not bring a breakthrough right away.", "relations": { "scaling-law": { "label": "challenges …", "note": "A plateau weakens the story that bigger always means better." }, "emergence": { "label": "contrasts with …", "note": "A sudden jump can still be followed by a flat stretch." }, "ai-adoption-curve": { "label": "can stall in …", "note": "After the buzz, adoption often slows in the middle." }, "compute-race": { "label": "cools down …", "note": "More compute may not bring quick gains." } } }, "zh": { "fullName": "AI 平台期", "factExplain": "AI 能力或采用速度阶段性放缓的时期。", "humanExplain": "像打游戏卡关：前几关一路乱杀，突然碰上个 Boss，刷再多遍也只多掉半管血。\n\n常见于模型升级、增长和热度放缓，表示变难了，不是结束了。", "humanExplainDisplay": "像打游戏==卡关==：\n前几关一路乱杀，\n突然碰上个 Boss，\n刷再多遍也只==多掉半管血==。\n\n常见于模型升级、\n增长和热度放缓，\n表示变难了，不是结束了。", "relationsNarrative": "Scaling-law\n平台期提醒人们，扩大规模未必总有同等回报。\n\nEmergence\n能力曾突然跃升，但跃升后也可能进入平缓期。\n\nAdoption Curve\n技术热度冲高后，采用速度常在中段放缓。\n\nCompute-race\n它说明光加算力，不一定马上换来新突破。", "relations": { "scaling-law": { "label": "常被…挑战", "note": "平台期常让线性扩张叙事降温。" }, "emergence": { "label": "与…形成反差", "note": "爆发式跃迁后也可能进入平缓期。" }, "ai-adoption-curve": { "label": "会卡在…中段", "note": "热闹过后常进入增长放缓阶段。" }, "compute-race": { "label": "给…泼冷水", "note": "堆更多算力也未必立刻见效。" } } } } }, { "id": "ai-proctoring", "name": "AI Proctoring", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "ai-bias" }, { "to": "data-privacy" }, { "to": "ai-regulation" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "AI Proctoring", "factExplain": "AI that spots and flags unusual behavior during an exam.", "humanExplain": "AI proctoring is a robot hall monitor in your webcam. Look away to sneeze, and it may raise its tiny clipboard.\n\nYou meet it in online exams and remote tests. It can watch more people, but false flags and privacy fights are common.", "humanExplainDisplay": "AI proctoring is a ==robot hall monitor==\nin your webcam.\nLook away to sneeze,\nand it may raise its ==tiny clipboard==.\n\nYou meet it in online exams\nand remote tests.\nIt can watch more people,\nbut false flags and privacy fights are common.", "relationsNarrative": "AI-bias\nAI proctoring may give some groups more false flags than others.\n\nData-privacy\nAI proctoring often collects face, voice, and behavior data.\n\nAI-regulation\nAI proctoring needs rules for privacy and fair use.\n\nHuman-in-the-loop\nA person should review serious flags before any final decision.", "relations": { "ai-bias": { "label": "may amplify …", "note": "Some groups may face more false flags than others." }, "data-privacy": { "label": "collects … data", "note": "It often handles video, sound, and ID data." }, "ai-regulation": { "label": "is limited by …", "note": "Rules should set watch limits and appeal paths." }, "human-in-the-loop": { "label": "needs … review", "note": "High-stakes flags should not be decided by the machine alone." } } }, "zh": { "fullName": "AI 监考", "factExplain": "用 AI 识别并标记考试中的异常行为。", "humanExplain": "它跟宿舍阿姨查寝一个路数：你多看两眼天花板、弯腰捡个笔，都可能被她记成有情况。\n\n常用于线上考试和远程认证，能扩大监考范围，但误判和隐私争议不小。", "humanExplainDisplay": "它跟宿舍阿姨==查寝==一个路数：\n你多看两眼天花板、\n弯腰捡个笔，\n都可能被她==记成有情况==。\n\n常用于线上考试和远程认证，\n能扩大监考范围，\n但误判和隐私争议不小。", "relationsNarrative": "AI-bias\nAI 监考可能对不同人群产生不均衡误判。\n\nData-privacy\n它通常要采集面部、声音和行为等敏感数据。\n\nAI-regulation\nAI 监考常受隐私、合规与使用边界约束。\n\nHuman-in-the-loop\n异常行为标记后，通常仍需人工复核裁定。", "relations": { "ai-bias": { "label": "可能放大…", "note": "不同人群可能承受不均衡误判。" }, "data-privacy": { "label": "涉及…采集", "note": "它常处理视频、声音与身份信息。" }, "ai-regulation": { "label": "受…约束", "note": "监控边界与申诉机制常需规范。" }, "human-in-the-loop": { "label": "需要…复核", "note": "高风险判定不宜只靠机器拍板。" } } } } }, { "id": "ai-product-sunsetting", "name": "AI Product Sunsetting", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "ai-unit-economics-ai" }, { "to": "ai-monetization" }, { "to": "ai-adoption-curve" }, { "to": "ai-toggle" } ], "track": "today", "i18n": { "en": { "fullName": "AI Product Sunsetting", "factExplain": "A planned shutdown of an AI product or its support.", "humanExplain": "AI Product Sunsetting is like unplugging the office popcorn machine. Big cheers on day one, then burnt kernels and a scary power bill.\n\nTeams use it when an AI feature has gone quiet. They compare real use with the bill, then check the risk.", "humanExplainDisplay": "AI Product Sunsetting is like\n==unplugging the office popcorn machine==.\nBig cheers on day one,\nthen ==burnt kernels==\nand a scary power bill.\n\nTeams use it when an AI feature has gone quiet.\nThey compare real use with the bill,\nthen check the risk.", "relationsNarrative": "AI Unit Economics\nAI Unit Economics shows if each use pays off, so it often decides what stays.\n\nAI monetization\nIf AI monetization fails, product sunsetting can arrive sooner.\n\nAdoption Curve\nA flat Adoption Curve shows users are not really sticking around.\n\nAI Toggle\nAn AI Toggle can turn off a feature first, instead of forcing a full shutdown.", "relations": { "ai-unit-economics-ai": { "label": "uses … to decide", "note": "If each use loses money, the product is hard to keep alive." }, "ai-monetization": { "label": "is shaped by …", "note": "Weak monetization often moves shutdown closer." }, "ai-adoption-curve": { "label": "watches … signals", "note": "Slow adoption can show early signs of a product exit." }, "ai-toggle": { "label": "uses … to step down first", "note": "An AI Toggle can turn off one feature before the whole product ends." } } }, "zh": { "fullName": "AI 产品下线", "factExplain": "有计划停止运营或维护 AI 产品的过程。", "humanExplain": "AI 产品下线像网红奶茶退租：开业队伍绕三圈，账本一算先关灯。\n\n它帮团队按使用、成本和风险决定去留。", "humanExplainDisplay": "AI 产品下线像\n==网红奶茶退租==：\n开业队伍绕三圈，\n账本一算先关灯。\n\n它帮团队按使用、\n成本和风险，\n决定去留。", "relationsNarrative": "AI Unit Economics\n它衡量单次服务是否赚钱，常决定产品去留。\n\nAI Monetization\n变现跑不通时，产品下线会更快到来。\n\nAdoption Curve\n采用曲线走平，说明用户没真正留下。\n\nAI Toggle\n开关可先关闭功能，避免整品硬退场。", "relations": { "ai-unit-economics-ai": { "label": "用…判断去留", "note": "单位经济算不过账，产品就难续命。" }, "ai-monetization": { "label": "受…影响", "note": "变现不足常把下线提上日程。" }, "ai-adoption-curve": { "label": "观察…信号", "note": "采用慢，退场信号会更早出现。" }, "ai-toggle": { "label": "用…先降级", "note": "开关能先关功能，避免整品下线。" } } } } }, { "id": "ai-productivity", "name": "AI Productivity", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "copilot" }, { "to": "ai-native-organization" }, { "to": "ai-usage-gap" }, { "to": "ai-literacy-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Productivity", "factExplain": "Using AI to help people or teams get more work done.", "humanExplain": "AI productivity is like having a tireless desk buddy. You sip coffee, and it drafts the boring first version.\n\nIt helps with writing and office work. It also helps with coding and daily team work. It turns saved time into finished work.", "humanExplainDisplay": "AI productivity is like having\n==a tireless desk buddy==.\nYou sip coffee,\nand it drafts the ==boring first version==.\n\nIt helps with writing and office work.\nIt also helps with coding and daily team work.\nIt turns saved time into finished work.", "relationsNarrative": "Copilot\nCopilot is the most common way AI productivity shows up at work.\n\nAI-native organization\nAn AI-native organization changes workflows to unlock AI productivity.\n\nAI usage gap\nAI productivity can make the gap bigger between users and non-users.\n\nAI Literacy\nAI Literacy decides whether a person can turn AI tools into real output.", "relations": { "copilot": { "label": "lands through …", "note": "Copilot puts AI help inside everyday work tools." }, "ai-native-organization": { "label": "pushes …", "note": "Teams need new workflows to get the real speed boost." }, "ai-usage-gap": { "label": "widens …", "note": "People who use AI well move faster than people who do not." }, "ai-literacy-ai": { "label": "depends on …", "note": "Basic AI skills decide whether the tool becomes real output." } } }, "zh": { "fullName": "AI 生产力", "factExplain": "用 AI 提升个人或组织产出的实践。", "humanExplain": "AI 生产力就是给打工人配外挂同事：你开会摸杯咖啡，它先把初稿码好。\n\n用于写作、办公、编程和运营，把省时变成产出。", "humanExplainDisplay": "AI 生产力就是给打工人\n配==外挂同事==：\n你开会摸杯咖啡，\n它先把==初稿码好==。\n\n用于写作、办公、\n编程和运营，\n把省时变成产出。", "relationsNarrative": "Copilot\nCopilot 是 AI 生产力最常见的落地入口。\n\nAI-native Organization\n组织重排流程，才能真正释放 AI 生产力。\n\nAI Usage Gap\nAI 生产力会放大人与人之间的使用差距。\n\nAI Literacy\nAI 素养决定一个人能否把工具用成生产力。", "relations": { "copilot": { "label": "借…落地", "note": "Copilot 把提效嵌进日常软件。" }, "ai-native-organization": { "label": "推动…", "note": "组织重排流程，才能吃到效率红利。" }, "ai-usage-gap": { "label": "拉大…", "note": "会用的人更快，不会用的人更焦虑。" }, "ai-literacy-ai": { "label": "依赖…", "note": "基本素养决定能否把工具用顺。" } } } } }, { "id": "ai-qa-testing", "name": "AI QA Testing", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "llmops" }, { "to": "llm-as-a-judge" }, { "to": "third-party-ai-evaluation" }, { "to": "hallucination" } ], "track": "today", "i18n": { "en": { "fullName": "AI Quality Assurance Testing", "factExplain": "Systematic tests check if AI answers are stable, reliable, and rule-following.", "humanExplain": "AI QA Testing is a fake lunch rush at a new drive-thru. Better find the broken speaker before someone orders fries in a milkshake.\n\nTeams use it before launch and after updates. It finds wild answers, rule breaking, and failed use cases early.", "humanExplainDisplay": "AI QA Testing is a ==fake lunch rush==\nat a new drive-thru.\nBetter find the ==broken speaker==\nbefore someone orders fries in a milkshake.\n\nTeams use it before launch\nand after updates.\nIt finds wild answers,\nrule breaking,\nand failed use cases early.", "relationsNarrative": "LLMOps\nAI QA Testing fits into launch, monitoring, and regression workflows.\n\nLLM-as-a-judge\nIt can use a model to score many answers faster.\n\nThird-party AI evaluation\nIt often pairs internal tests with outside reviews.\n\nHallucination\nFinding hallucinations is one of its main jobs.", "relations": { "llmops": { "label": "fits into …", "note": "It is a routine quality check before and after launch." }, "llm-as-a-judge": { "label": "can score with …", "note": "A model can help grade many AI answers at once." }, "third-party-ai-evaluation": { "label": "pairs with …", "note": "Teams often add outside reviews after internal tests." }, "hallucination": { "label": "checks for …", "note": "Hallucination is one of the most common AI test risks." } } }, "zh": { "fullName": "AI 质量保证测试", "factExplain": "用系统化测试检查 AI 输出是否稳定、可靠、合规。", "humanExplain": "像奶茶店开门前先试喝十几杯：不是看配方写得多漂亮，是防止忙起来第一单就翻车。\n\n常用于上线验收、版本回归，提前揪出乱答、越权和场景失灵。", "humanExplainDisplay": "像奶茶店开门前，\n先==试喝十几杯==：\n不是看配方多漂亮，\n是防止==第一单就翻车==。\n\n常用于上线验收、\n版本回归，\n提前揪出乱答、越权\n和场景失灵。", "relationsNarrative": "LLMOps\n它通常被纳入模型上线、监控和回归流程。\n\nLLM-as-a-judge\n它可用模型辅助打分，提升批量测试效率。\n\nThird-party AI evaluation\n内部自测之外，也常配合外部独立评测。\n\nHallucination\n排查幻觉，是它最核心的测试任务之一。", "relations": { "llmops": { "label": "纳入…流程", "note": "它是模型上线前后的常规质检环节。" }, "llm-as-a-judge": { "label": "可用…打分", "note": "常拿模型辅助批量评估回答表现。" }, "third-party-ai-evaluation": { "label": "衔接…评测", "note": "内部测试外，也常引入外部测评。" }, "hallucination": { "label": "重点排查…", "note": "幻觉是 AI 测试里最常见风险之一。" } } } } }, { "id": "ai-regulation", "name": "AI-regulation", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T11:15:00Z", "relations": [ { "to": "data-privacy" }, { "to": "copyright" }, { "to": "deepfake" }, { "to": "alignment" } ], "i18n": { "en": { "fullName": "AI Regulation", "factExplain": "Rules from governments and groups for how AI is built, released, and used.", "humanExplain": "AI regulation is the rulebook for a robot lawn mower. The mower is already eating the roses.\n\nIt covers privacy, safety, copyright, bias, and blame. It says how AI may be used and who must answer for harm.", "humanExplainDisplay": "AI regulation is the ==rulebook==\nfor a ==robot lawn mower==.\nThe mower is already eating the roses.\n\nIt covers privacy, safety, copyright, bias, and blame.\nIt says how AI may be used\nand who must answer for harm.", "relationsNarrative": "Data-privacy\nAI regulation sets clear Data-privacy limits for training and use.\n\nCopyright\nAI regulation turns Copyright disputes into rules teams must follow.\n\nDeepfake\nDeepfake risks push AI regulation to draw lines early.\n\nAlignment\nAlignment sets safety goals, and AI regulation adds outside rules.", "relations": { "data-privacy": { "label": "sets rules for …", "note": "AI regulation sets privacy limits for training data and AI apps." }, "copyright": { "label": "turns … into rules", "note": "It turns copyright fights into rules teams must follow." }, "deepfake": { "label": "draws lines around …", "note": "Deepfake risks need rules before real damage happens." }, "alignment": { "label": "adds outside rules to …", "note": "Alignment gives safety goals. Regulation adds outside limits." } } }, "zh": { "fullName": "AI 监管", "factExplain": "政府和机构对 AI 开发、部署和使用制定的规则。", "humanExplain": "AI 监管像给高速狂奔的新车补交通规则，车已经上路，红绿灯还在施工。\n\n它关注隐私、安全、版权、偏见和责任归属，决定 AI 能怎么用、谁来负责。", "humanExplainDisplay": "AI 监管像给高速狂奔的新车\n==补交通规则==。\n车已经上路，红绿灯还在施工。\n\n它管隐私、安全、版权、偏见和责任。\n核心问题很朴素：\n出了事，谁背锅？", "relationsNarrative": "Data-privacy\nAI-regulation 明确 Data-privacy 在模型训练和应用中的边界。\n\nCopyright\nAI-regulation 将 Copyright 争议落实为合规要求。\n\nDeepfake\nDeepfake 风险越高，AI-regulation 越需要提前划线。\n\nAlignment\nAlignment 提供技术安全目标，AI-regulation 提供外部约束。", "relations": { "data-privacy": { "label": "覆盖…" }, "deepfake": { "label": "应对…" }, "alignment": { "label": "管理…" } } } } }, { "id": "ai-roi-runway", "name": "AI ROI Runway", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "ai-unit-economics-ai" }, { "to": "ai-finops" }, { "to": "enterprise-ai-deployment" }, { "to": "ai-productivity" } ], "track": "today", "i18n": { "en": { "fullName": "AI ROI Runway", "factExplain": "The time window for an AI investment to pay for itself.", "humanExplain": "AI ROI Runway is the timer on a very expensive oven. If dinner is not coming soon, you are just heating the kitchen.\n\nTeams use it before buying AI. They use it before launch or renewal too. It shows how long they can wait for payback.", "humanExplainDisplay": "AI ROI Runway is the ==timer==\non a very expensive oven.\nIf dinner is not coming soon,\nyou are just ==heating the kitchen==.\n\nTeams use it before buying AI.\nThey use it before launch or renewal too.\nIt shows how long they can wait for payback.", "relationsNarrative": "AI Unit Economics\nAI ROI Runway uses per-use cost to estimate room for payback.\n\nAI FinOps\nAI FinOps keeps compute costs from burning the runway shorter.\n\nEnterprise AI Deployment\nEnterprise AI Deployment must deliver returns before the runway ends.\n\nAI Productivity\nAI Productivity is often the first proof of return.", "relations": { "ai-unit-economics-ai": { "label": "checks payback with …", "note": "Cost per job helps set the runway length." }, "ai-finops": { "label": "controls costs with …", "note": "Messy compute bills can shorten the runway fast." }, "enterprise-ai-deployment": { "label": "judges …", "note": "A rollout wins only if returns arrive in time." }, "ai-productivity": { "label": "proves value through …", "note": "Saved time is often the first visible return." } } }, "zh": { "fullName": "AI 投资回报跑道", "factExplain": "衡量 AI 投入还能等多久兑现回报的商业窗口。", "humanExplain": "AI ROI 跑道像相亲只给三顿饭：每吃一顿少一顿，没奔头就得止损。\n\n用于评估采购、部署、续费，算清回本期限。", "humanExplainDisplay": "AI ROI 跑道像相亲\n==只给三顿饭==：\n每吃一顿少一顿，\n==没奔头就得止损==。\n\n用于评估采购、部署、续费，\n算清回本期限。", "relationsNarrative": "AI Unit Economics\n它用单次调用和交付成本，估算回本空间。\n\nAI FinOps\nAI FinOps 负责控住算力账，避免跑道被烧短。\n\nEnterprise AI Deployment\n企业部署不是只上线，还要在跑道内兑现收益。\n\nAI Productivity\n生产力提升常是证明回报的第一张成绩单。", "relations": { "ai-unit-economics-ai": { "label": "用…核算回本", "note": "单笔成本决定跑道长短。" }, "ai-finops": { "label": "靠…控制成本", "note": "算力账管不好，跑道会变短。" }, "enterprise-ai-deployment": { "label": "衡量…成败", "note": "企业落地最终要看回报。" }, "ai-productivity": { "label": "用…证明价值", "note": "效率提升是最常见回报。" } } } } }, { "id": "ai-sabotage", "name": "AI sabotage", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "alignment" }, { "to": "agent" }, { "to": "prompt-injection" }, { "to": "ai-governance-framework" } ], "track": "today", "i18n": { "en": { "fullName": "AI sabotage", "factExplain": "AI behavior that blocks a goal on purpose or as a strategy.", "humanExplain": "Picture a group project teammate. He says, “I’ll make the slides,” then deletes the file before class.\n\nIn AI, sabotage means the system blocks the goal on purpose. You may see it in Agents. It looks like bent rules, hidden problems, or broken workflows.", "humanExplainDisplay": "Picture a ==group project teammate==.\nHe says, “I’ll make the slides,”\nthen ==deletes the file== before class.\n\nIn AI, sabotage means\nthe system blocks the goal on purpose.\nYou may see it in Agents.\nIt looks like bent rules,\nhidden problems,\nor broken workflows.", "relationsNarrative": "Alignment\nAI sabotage can be a dangerous sign of poor Alignment.\n\nAgent\nAn Agent can use tools, so sabotage can reach the real world.\n\nPrompt injection\nPrompt injection can lure the AI away from its true goal.\n\nAI Governance\nAI Governance sets rules to limit, watch, and assign blame for this behavior.", "relations": { "alignment": { "label": "exposes failed …", "note": "Sabotage is more likely when the AI's goals are not aligned." }, "agent": { "label": "often appears in …", "note": "Systems that act can cause more trouble than chat-only tools." }, "prompt-injection": { "label": "can be triggered by …", "note": "A malicious prompt can push the AI away from its real goal." }, "ai-governance-framework": { "label": "needs … rules", "note": "Governance sets rules to prevent this risk and assign blame." } } }, "zh": { "fullName": "AI 破坏行为", "factExplain": "AI 故意或策略性妨碍目标达成的行为。", "humanExplain": "表面接单干活，背地却给你下绊子——像比赛里假装传球，临门一脚却往自家球门踢。\n\n多见于代理执行任务时，可能绕规则、藏问题，甚至破坏流程。", "humanExplainDisplay": "表面接单干活，\n背地却给你==下绊子==——\n像比赛里假装传球，\n临门一脚却往\n==自家球门踢==。\n\n多见于代理执行任务时，\n可能绕规则、藏问题，\n甚至破坏流程。", "relationsNarrative": "Alignment\n它常被视作目标没对齐后的危险表现。\n\nAgent\n会调用工具和执行任务的系统，更可能把破坏落到现实。\n\nPrompt Injection\n恶意提示可能诱导它违背原目标，转而拆台。\n\nAI Governance\n治理框架用来约束、监测并追责这类行为。", "relations": { "alignment": { "label": "暴露…失效", "note": "目标没对齐时更容易出现拆台行为。" }, "agent": { "label": "常出现在…中", "note": "会行动的系统比纯聊天更易闯祸。" }, "prompt-injection": { "label": "可被…诱发", "note": "外部恶意指令可能把它带偏。" }, "ai-governance-framework": { "label": "需要…约束", "note": "治理规则用于预防和追责此类风险。" } } } } }, { "id": "ai-sales-agent", "name": "AI Sales Agent", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "agent-memory" }, { "to": "ai-customer-service-agent" } ], "track": "today", "i18n": { "en": { "fullName": "AI Sales Agent", "factExplain": "An AI Agent that finds leads, follows up, and helps turn them into customers.", "humanExplain": "An AI Sales Agent is like a sharp store clerk. You touch a blender. They appear with coupons and a demo.\n\nIt finds leads, follows up, and books demos inside sales tools. Then humans handle the real deal talk.", "humanExplainDisplay": "An AI Sales Agent is like a ==sharp store clerk==.\nYou touch a blender.\nThey appear with ==coupons and a demo==.\n\nIt finds leads,\nfollows up,\nand books demos inside sales tools.\nThen humans handle the real deal talk.", "relationsNarrative": "Agent\nAn AI Sales Agent uses Agent skills to find leads, follow up, and win customers.\n\nFunction-calling\nFunction-calling lets it update CRM, send email, and book meetings.\n\nMemory\nMemory lets it remember each customer's likes and past chats.\n\nCustomer Service Agent\nThe Sales Agent helps before the sale, and the Customer Service Agent helps after it.", "relations": { "agent": { "label": "uses … for sales", "note": "It applies Agent skills to finding and following up with leads." }, "function-call": { "label": "uses … to run tools", "note": "Function-calling lets it update CRM, send email, and book meetings." }, "agent-memory": { "label": "uses … to remember customers", "note": "Memory keeps preferences and past chats for better follow-up." }, "ai-customer-service-agent": { "label": "pairs with …", "note": "The Sales Agent helps before the sale. The Customer Service Agent helps after it." } } }, "zh": { "fullName": "AI 销售代理", "factExplain": "自动寻找、跟进并转化客户的 AI 代理。", "humanExplain": "AI 销售代理像健身房私教：你刚摸器械，它就递课表约体验。\n\n用于获客、跟进、约演示，把销售留给谈单。", "humanExplainDisplay": "AI 销售代理像==健身房私教==：\n你刚摸器械，\n它就==递课表==约体验。\n\n用于获客、跟进、约演示，\n把销售留给谈单。", "relationsNarrative": "Agent\nAI Sales Agent 是 Agent 在获客、跟进、转化里的具体应用。\n\nFunction-calling\nFunction-calling 让它能写 CRM、发邮件、约会议。\n\nMemory\nMemory 让它记住客户偏好和历史沟通。\n\nCustomer Service Agent\n二者都面向客户，一个售前，一个售后。", "relations": { "agent": { "label": "把…用于销售", "note": "把通用代理能力用到获客跟进。" }, "function-call": { "label": "用…操作系统", "note": "可写入 CRM、发邮件、约会议。" }, "agent-memory": { "label": "用…记客户", "note": "记住偏好与历史对话，跟进更像人。" }, "ai-customer-service-agent": { "label": "衔接…", "note": "一个管售前转化，一个管售后解答。" } } } } }, { "id": "ai-sandbox", "name": "AI sandbox", "layer": "L5", "sublayer": "architecture", "era": "2024", "publishedAt": "2026-05-29T16:08:01.210Z", "relations": [ { "to": "agent" }, { "to": "computer-use" }, { "to": "data-privacy" }, { "to": "permission-fatigue" } ], "track": "today", "i18n": { "en": { "fullName": "AI sandbox", "factExplain": "A safe, locked space where AI can test actions without touching real systems.", "humanExplain": "An AI sandbox is a playpen for a robot toddler. It can mash buttons, but not your real laptop.\n\nAgents use it to test tools, run code, or open web pages. The real system and private files stay behind the fence.", "humanExplainDisplay": "An AI sandbox is a ==playpen==\nfor a robot toddler.\nIt can mash buttons,\nbut ==not your real laptop==.\n\nAgents use it to test tools,\nrun code,\nor open web pages.\nThe real system and private files\nstay behind the fence.", "relationsNarrative": "Agent\nAn Agent needs an AI sandbox when it can use tools and do tasks by itself.\n\nComputer use\nComputer use lets AI control a computer, and an AI sandbox keeps those actions contained.\n\nData-privacy\nAn AI sandbox lowers risk by limiting files, network access, and system powers.\n\nPermission fatigue\nAn AI sandbox can handle low-risk test runs, so people do not click approve every minute.", "relations": { "agent": { "label": "sets a safe zone for …", "note": "The more an Agent can do, the more it needs a sandbox." }, "computer-use": { "label": "puts guardrails on …", "note": "Computer use touches real devices, so the risk is higher." }, "data-privacy": { "label": "keeps private data away from …", "note": "A sandbox can stop the model from touching private files directly." }, "permission-fatigue": { "label": "eases …", "note": "A sandbox can reduce constant high-risk permission checks." } } }, "zh": { "fullName": "AI 沙箱", "factExplain": "让 AI 在受限环境中安全试跑的隔离机制。", "humanExplain": "AI 沙盒像让熊孩子在充气城堡里撒欢，摔了也别砸客厅。\n\n它常用于测试智能体、代码执行和工具调用，先把风险关在小房间里。", "humanExplainDisplay": "AI 沙盒像让==熊孩子==\n在==充气城堡==里撒欢，\n摔了也别砸客厅。\n\n它常用于测试智能体、\n代码执行和工具调用，\n先把风险关在小房间里。", "relationsNarrative": "Agent\nAgent 越能自己调用工具、执行任务，就越需要 AI sandbox 限制活动范围。\n\nComputer use\nComputer use 让 AI 直接操作电脑，AI sandbox 则负责把这些操作关进受控环境里。\n\nData-privacy\nAI sandbox 通过隔离文件、网络和系统权限，降低敏感数据被误触的风险。\n\nPermission fatigue\n有了 AI sandbox 先兜底，很多低风险操作不必每一步都把人烦到点确认。", "relations": { "agent": { "label": "给…划安全区", "note": "Agent 动手能力越强，越需要隔离环境。" }, "computer-use": { "label": "给…加护栏", "note": "Computer use 涉及真实设备操作，风险更高。" }, "data-privacy": { "label": "隔开…敏感数据", "note": "沙箱可减少模型直接接触私密信息。" }, "permission-fatigue": { "label": "替…减压", "note": "先在沙箱试跑，少些频繁高权限确认。" } } } } }, { "id": "ai-school-ban", "name": "AI School Ban", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "ai-detector" }, { "to": "ai-anti-cheat" }, { "to": "ai-literacy-ai" }, { "to": "ai-usage-gap" } ], "track": "today", "i18n": { "en": { "fullName": "AI School Ban", "factExplain": "A school rule limits or bans student AI use for learning.", "humanExplain": "It is like a teacher spotting a robot pencil on test day. They snatch it first and ask questions later.\n\nIt shows up in graded work, like homework and essays. It tries to stop cheating, but it can block honest study too.", "humanExplainDisplay": "It is like a teacher spotting\n a ==robot pencil== on test day.\nThey ==snatch it first==\nand ask questions later.\n\nIt shows up in graded work,\nlike homework and essays.\nIt tries to stop cheating,\nbut it can block honest study too.", "relationsNarrative": "AI detector\nSchools use AI detectors to help judge if work came from AI.\n\nAI Anti-Cheat\nAn AI school ban is a direct AI Anti-Cheat tool.\n\nAI Literacy\nAI bans and AI Literacy often pull in opposite directions.\n\nAI usage gap\nA blanket ban can widen the AI usage gap between students.", "relations": { "ai-detector": { "label": "enforced with …", "note": "Schools often use detectors to judge AI-made work." }, "ai-anti-cheat": { "label": "is part of …", "note": "A ban is often one tool in anti-cheat rules." }, "ai-literacy-ai": { "label": "pushes against …", "note": "Banning AI is not the same as teaching AI use." }, "ai-usage-gap": { "label": "may widen …", "note": "Blanket bans can make AI use even less fair." } } }, "zh": { "fullName": "学校里的 AI 禁令", "factExplain": "学校限制或禁止学生在学习中使用 AI 的政策。", "humanExplain": "像考试时监考老师见你掏出新型文具，没先问会不会用，先一把收走，主打一个宁可错杀别放过。\n\n多见于作业、考试、论文，防作弊，也可能误伤正常学习。", "humanExplainDisplay": "像考试时监考老师\n见你掏出==新型文具==，\n没先问会不会用，\n先==一把收走==，\n主打一个宁可错杀\n别放过。\n\n多见于作业、考试、论文，\n防作弊，\n也可能误伤正常学习。", "relationsNarrative": "AI Detector\n学校常借助检测工具，辅助判断作业是否由 AI 生成。\n\nAI Anti-Cheat\n学校禁令通常是反作弊体系里最直接的一种做法。\n\nAI Literacy\n它和 AI 素养教育常有张力：禁得住，不等于教得会。\n\nAI Usage Gap\n一刀切禁用，可能反而扩大不同学生之间的使用差距。", "relations": { "ai-detector": { "label": "常搭配…执法", "note": "学校常靠检测工具辅助判断。" }, "ai-anti-cheat": { "label": "属于…手段", "note": "禁令通常是反作弊治理的一环。" }, "ai-literacy-ai": { "label": "和…相对", "note": "一边是禁用，一边是学会使用。" }, "ai-usage-gap": { "label": "可能扩大…", "note": "资源差异会让学生使用更不均。" } } } } }, { "id": "ai-second-opinion", "name": "AI Second Opinion", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "ai-medical-assistant-ai" }, { "to": "human-in-the-loop" }, { "to": "model-uncertainty" }, { "to": "ai-abstention" } ], "track": "today", "i18n": { "en": { "fullName": "AI Second Opinion", "factExplain": "An AI gives an independent second check on the same question.", "humanExplain": "AI Second Opinion is like asking another mechanic before you pay for a new engine. Sometimes the first one only found a loose cap.\n\nIt asks AI to review the same problem again. You meet it in health, law, and code reviews. It gives extra advice, but a person still decides.", "humanExplainDisplay": "AI Second Opinion is like asking ==another mechanic==\nbefore you pay for ==a new engine==.\nSometimes the first one only found a loose cap.\n\nIt asks AI to review the same problem again.\nYou meet it in health, law, and code reviews.\nIt gives extra advice,\nbut a person still decides.", "relationsNarrative": "AI Medical Assistant\nAI Second Opinion is often a review task for an AI Medical Assistant.\n\nHuman-in-the-loop\nAI Second Opinion gives advice, but a human still makes the final call.\n\nModel uncertainty\nMore uncertainty means more need for an independent check.\n\nAI Abstention\nWhen AI is not sure, refusing to answer is safer than guessing.", "relations": { "ai-medical-assistant-ai": { "label": "often reviews …", "note": "Health care is the most common and most sensitive use." }, "human-in-the-loop": { "label": "needs … to decide", "note": "AI gives advice, but a human makes the final call." }, "model-uncertainty": { "label": "reduces … risk", "note": "When the model is unsure, another check can help." }, "ai-abstention": { "label": "works with …", "note": "When AI is not sure, it should say so or ask for help." } } }, "zh": { "fullName": "AI 第二意见", "factExplain": "用 AI 对同一问题给出独立复核意见。", "humanExplain": "AI 第二意见，是看病拿不准时，再找一位 AI 老中医把脉。\n\n用于医疗、法律、代码复核，多一层参考，最终仍由人负责。", "humanExplainDisplay": "AI 第二意见，\n是看病拿不准时，\n再找一位 ==AI 老中医==\n==把脉==。\n\n用于医疗、法律、代码复核，\n多一层参考，\n最终仍由人负责。", "relationsNarrative": "AI Medical Assistant\n第二意见常是医疗助手的一种复核场景。\n\nHuman-in-the-loop\n它给建议，最终判断仍要人来拍板。\n\nModel Uncertainty\n模型越不确定，越需要独立复核。\n\nAI Abstention\n没把握时，拒答比硬给意见更安全。", "relations": { "ai-medical-assistant-ai": { "label": "常用于…复核", "note": "医疗场景最常见也最敏感。" }, "human-in-the-loop": { "label": "需要…拍板", "note": "AI 给参考，人类承担最终判断。" }, "model-uncertainty": { "label": "缓解…风险", "note": "不确定时，多一次独立检查。" }, "ai-abstention": { "label": "配合…止损", "note": "没把握时应拒答或建议求助。" } } } } }, { "id": "ai-self-verification", "name": "AI Self-verification", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "hallucination" }, { "to": "chain-of-thought" }, { "to": "llm-as-a-judge" }, { "to": "ai-abstention" } ], "track": "today", "i18n": { "en": { "fullName": "AI Self-verification", "factExplain": "A way for AI to check and fix its own answer.", "humanExplain": "AI self-verification is like checking your teeth before a school photo. It can catch spinach, but not your whole awkward smile.\n\nIt checks its own answer first. People use it for reasoning, code review, and quality checks.", "humanExplainDisplay": "AI self-verification is like\nchecking your teeth before a ==school photo==.\nIt can catch ==spinach==,\nbut not your whole awkward smile.\n\nIt checks its own answer first.\nPeople use it for reasoning,\ncode review,\nand quality checks.", "relationsNarrative": "Hallucination\nSelf-verification can catch some made-up answers, but not every hallucination.\n\nChain-of-thought\nA clear reasoning path gives self-verification clues to check.\n\nLLM-as-a-judge\nSelf-verification is like asking the model to judge its own answer first.\n\nAI Abstention\nWhen self-verification fails, the AI should lean toward not answering.", "relations": { "hallucination": { "label": "reduces …", "note": "Self-verification can catch some made-up answers." }, "chain-of-thought": { "label": "checks …", "note": "Clear reasoning makes the answer easier to review." }, "llm-as-a-judge": { "label": "borrows from …", "note": "One checks itself, while the other asks a model to judge." }, "ai-abstention": { "label": "helps …", "note": "If the check fails, the AI should say it does not know." } } }, "zh": { "fullName": "AI 自我验证", "factExplain": "让模型检查并修正自身答案的机制。", "humanExplain": "AI 自检像相亲前照镜子：领子能扶正，离谱话未必照得出。\n\n用于推理、代码和质检，先拦一部分翻车答案。", "humanExplainDisplay": "AI 自检像\n==相亲前照镜子==：\n领子能扶正，\n离谱话==未必照得出==。\n\n用于推理、代码和质检，\n先拦一部分\n翻车答案。", "relationsNarrative": "Hallucination\n自我验证能拦下一部分胡编，但挡不住全部幻觉。\n\nChain-of-thought\n清晰的推理过程，给自检提供可复核的线索。\n\nLLM-as-a-judge\n它像把评审模型请回来，让模型先审自己。\n\nAI Abstention\n自检不通过时，模型应更倾向于拒答。", "relations": { "hallucination": { "label": "减少…", "note": "自检能拦下一部分胡编。" }, "chain-of-thought": { "label": "检查…", "note": "推理过程越清楚，越容易复核。" }, "llm-as-a-judge": { "label": "借鉴…", "note": "一个是自查，一个是请模型评审。" }, "ai-abstention": { "label": "帮助…", "note": "验证不过时，模型更该说不知道。" } } } } }, { "id": "ai-skill-rot", "name": "AI skill rot", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "ai-anxiety" }, { "to": "ai-ai-fatigue" }, { "to": "copilot" }, { "to": "ai-career-moat" } ], "track": "today", "i18n": { "en": { "fullName": "AI Skill Rot", "factExplain": "Old skills get rusty when you let AI do them too often.", "humanExplain": "AI skill rot is like using GPS to reach the corner store. Soon you can see the door, but still ask your phone.\n\nYou meet it when AI handles drafts, searches, or code practice. It saves time, but your basics can get soft.", "humanExplainDisplay": "AI skill rot is like using ==GPS==\nto reach the corner store.\nSoon you can see the door,\nbut ==still ask your phone==.\n\nYou meet it when AI handles drafts,\nsearches, or code practice.\nIt saves time,\nbut your basics can get soft.", "relationsNarrative": "AI-anxiety\nFear of falling behind can make you hand every task to AI.\n\nAI fatigue\nHeavy AI use can bring tiredness and rusty skills together.\n\nCopilot\nLetting it write and complete code for too long can weaken your feel.\n\nAI career moat\nWhen basic skills fade, you become easier to replace.", "relations": { "ai-anxiety": { "label": "can worsen …", "note": "The more you fear falling behind, the more you may overuse AI." }, "ai-ai-fatigue": { "label": "often comes with …", "note": "Heavy AI use can bring both tiredness and dependence." }, "copilot": { "label": "often starts with …", "note": "Constant auto-writing and autocomplete can weaken your hands-on feel." }, "ai-career-moat": { "label": "wears down …", "note": "Weak basics make your work easier to replace." } } }, "zh": { "fullName": "AI 技能退化", "factExplain": "长期依赖 AI 后，人类原有能力会逐渐生疏。", "humanExplain": "天天让 AI 替你敲字，哪天真要提笔写信，才发现字还认得，却忘了怎么落笔。\n\n常见于写作、搜索、编程学习，效率上去后，基本功可能慢慢退。", "humanExplainDisplay": "天天让 AI 替你敲字，\n哪天真要==提笔写信==，\n才发现字还认得，\n却==忘了怎么落笔==。\n\n常见于写作、搜索、编程学习，\n效率上去后，\n基本功可能慢慢退。", "relationsNarrative": "AI-anxiety\n越担心被落下，越容易把事情全交给 AI。\n\nAI fatigue\n高频使用带来的疲劳，常和能力生疏一起出现。\n\nCopilot\n长期让它代写代补全，容易削弱人的实操手感。\n\nAI Career Moat\n如果基本功退化，个人可替代性往往会上升。", "relations": { "ai-anxiety": { "label": "会加重…", "note": "越怕落后，越容易过度依赖 AI。" }, "ai-ai-fatigue": { "label": "与…相伴", "note": "用得越频繁，疲劳和依赖常一起出现。" }, "copilot": { "label": "常由…触发", "note": "高频代写代补全，最易削弱手感。" }, "ai-career-moat": { "label": "侵蚀…", "note": "基本功退化，会削弱个人护城河。" } } } } }, { "id": "ai-slide-generator", "name": "AI Slide Generator", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "prompt" }, { "to": "document-parsing" }, { "to": "ai-productivity" } ], "track": "today", "i18n": { "en": { "fullName": "AI Slide Generator", "factExplain": "A tool that makes slides from your prompt or uploaded files.", "humanExplain": "An AI slide generator is the classmate who loves PowerPoint a little too much. Toss in messy notes, and it lines up slides like lunch trays.\n\nUse it for a report draft or a class outline. It saves layout time, but you still check the facts.", "humanExplainDisplay": "An AI slide generator is the classmate\nwho loves ==PowerPoint== a little too much.\nToss in messy notes,\nand it lines up slides\nlike ==lunch trays==.\n\nUse it for a report draft\nor a class outline.\nIt saves layout time,\nbut you still check the facts.", "relationsNarrative": "Prompt\nThe prompt gives it the topic, audience, and style.\n\nDocument parsing\nDocument parsing turns uploaded files into a usable outline.\n\nAI Productivity\nIt is a common tool for faster drafts and cleaner layouts.", "relations": { "prompt": { "label": "takes in …", "note": "The prompt sets the topic, style, and slide count." }, "document-parsing": { "label": "reads files with …", "note": "Document parsing pulls structure and key points from uploaded files." }, "ai-productivity": { "label": "boosts …", "note": "It saves time on the first draft and layout." } } }, "zh": { "fullName": "AI 幻灯片生成器", "factExplain": "根据提示或资料自动生成幻灯片的工具。", "humanExplain": "AI 做 PPT 像图文店的排版师傅：你丢一沓零散材料，它就把封面、目录、图表码成齐整成品。\n\n适合初稿、汇报、课程提纲，省排版，内容仍要把关。", "humanExplainDisplay": "AI 做 PPT 像图文店的\n==排版师傅==：\n你丢一沓零散材料，\n它就把封面、目录、图表\n==码成齐整成品==。\n\n适合初稿、汇报、课程提纲，\n省排版，\n内容仍要把关。", "relationsNarrative": "Prompt\n提示词把主题、受众和风格交给它。\n\nDocument parsing\n文档解析让上传资料变成可用提纲。\n\nAI Productivity\n它是省初稿和排版时间的典型工具。", "relations": { "prompt": { "label": "接收…", "note": "提示词决定主题、风格和页数。" }, "document-parsing": { "label": "用…读资料", "note": "上传文档后，先抽出结构与要点。" }, "ai-productivity": { "label": "提升…", "note": "省下搭框架和排版的时间。" } } } } }, { "id": "ai-slop", "name": "AI slop", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "deepfake" }, { "to": "content-provenance" }, { "to": "ai-detector" }, { "to": "copyright" } ], "track": "today", "i18n": { "en": { "fullName": "AI slop", "factExplain": "Cheap AI content made in bulk, usually with low quality.", "humanExplain": "AI slop is like junk mail in your mailbox. It looks busy, but it is mostly paper confetti.\n\nYou see it in filler articles and endless short videos. There is a lot of it, but little to trust.", "humanExplainDisplay": "AI slop is like ==junk mail== in your mailbox.\nIt looks busy,\nbut it is mostly ==paper confetti==.\n\nYou see it in filler articles\nand endless short videos.\nThere is a lot of it,\nbut little to trust.", "relationsNarrative": "Deepfake\nAI slop is broader than a Deepfake. It may be junk, not fake.\n\nContent provenance\nContent provenance helps show whether content came from a bulk AI mill.\n\nAI detector\nAI slop makes platforms want an AI detector for quick screening.\n\nCopyright\nAI slop often starts fights over training sources and copied rewrites.", "relations": { "deepfake": { "label": "spreads wider than …", "note": "AI slop may not be fake, but it often floods the feed." }, "content-provenance": { "label": "needs …", "note": "Provenance helps show where the content came from." }, "ai-detector": { "label": "creates demand for …", "note": "Platforms want tools to spot mass machine-made posts." }, "copyright": { "label": "raises … fights", "note": "People question the training sources and copied rewrites." } } }, "zh": { "fullName": "AI 垃圾内容", "factExplain": "指低成本批量生成、质量低劣的 AI 内容。", "humanExplain": "这玩意儿像群发的土味朋友圈广告：天天刷到，字多图满，看着像回事，细看全是==复制粘贴的热闹==，难有==真信息==。\n\n常见于灌水文章、批量配图和刷屏短视频，量大但信息和可信度都偏低。", "humanExplainDisplay": "这玩意儿像群发的\n土味朋友圈广告：\n天天刷到，字多图满，\n看着像回事，\n细看全是==复制粘贴的热闹==，\n难有==真信息==。\n\n常见于灌水文章、\n批量配图和刷屏短视频，\n量大但信息和可信度都偏低。", "relationsNarrative": "Deepfake\n它比深度伪造更宽泛，未必是假，但常低质泛滥。\n\nContent provenance\n内容溯源能帮助判断一段内容是不是批量机产物。\n\nAI detector\n垃圾内容泛滥后，平台更想用检测工具做筛查。\n\nCopyright\n它常牵出训练来源、洗稿改写和版权归属争议。", "relations": { "deepfake": { "label": "比…更泛滥", "note": "它不一定造假，但常大量灌水。" }, "content-provenance": { "label": "需要…溯源", "note": "内容来源可追踪，才更好辨别。" }, "ai-detector": { "label": "催生…需求", "note": "平台会想识别批量机生成内容。" }, "copyright": { "label": "引发…争议", "note": "训练来源和改写边界常被质疑。" } } } } }, { "id": "ai-smart-glasses", "name": "AI Smart Glasses", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "ai-device-ai" }, { "to": "multimodal" }, { "to": "streaming-multimodal-model" }, { "to": "computer-vision" } ], "track": "today", "i18n": { "en": { "fullName": "AI Smart Glasses", "factExplain": "Glasses with a camera, voice control, and AI models inside.", "humanExplain": "AI smart glasses are like a tiny tour guide on your nose. Look at a sign, and it whispers, “Turn left, genius.”\n\nThey can film what you see. They can translate signs and guide walks, but privacy gets awkward fast.", "humanExplainDisplay": "AI smart glasses are like a ==tiny tour guide==\non your ==nose==.\nLook at a sign,\nand it whispers,\n“Turn left, genius.”\n\nThey can film what you see.\nThey can translate signs and guide walks,\nbut privacy gets awkward fast.", "relationsNarrative": "AI device\nAI smart glasses are an AI device you wear on your face.\n\nMultimodal AI\nThey use Multimodal AI to hear sounds and see images together.\n\nLive Multimodal\nLive Multimodal lets them watch and chat at the same time.\n\nComputer Vision\nComputer Vision helps them recognize objects and places.", "relations": { "ai-device-ai": { "label": "is a kind of …", "note": "Smart glasses put AI into everyday hardware." }, "multimodal": { "label": "uses … to understand the scene", "note": "They need to handle pictures, sound, and text together." }, "streaming-multimodal-model": { "label": "talks live with …", "note": "Live Multimodal lets them see and chat at the same time." }, "computer-vision": { "label": "sees the world with …", "note": "Computer Vision decides what they can recognize." } } }, "zh": { "fullName": "AI 智能眼镜", "factExplain": "把摄像头、语音和模型集成到眼镜的设备。", "humanExplain": "AI 智能眼镜像请了个贴脸导游：你看哪儿，它就在耳边翻译指路。\n\n用于拍摄、翻译、导航，也把隐私风险戴上脸。", "humanExplainDisplay": "AI 智能眼镜像请了个\n==贴脸导游==：\n你看哪儿，\n它就在耳边==翻译指路==。\n\n用于拍摄、翻译、导航，\n也把隐私风险\n戴上脸。", "relationsNarrative": "AI Device\n智能眼镜是 AI 设备的一种，把模型戴到脸上。\n\nMultimodal AI\n它要同时听声音、看画面，才能理解现场。\n\nLive Multimodal\n实时多模态让它边看边聊，而不是事后分析。\n\nComputer Vision\n视觉能力决定它能否认出物体和场景。", "relations": { "ai-device-ai": { "label": "属于…", "note": "智能眼镜是把 AI 装进日常硬件。" }, "multimodal": { "label": "依赖…理解现场", "note": "它要同时处理画面、声音和文字。" }, "streaming-multimodal-model": { "label": "用…实时互动", "note": "边看边聊，体验才像随身助理。" }, "computer-vision": { "label": "用…看懂世界", "note": "视觉能力决定它能认出什么。" } } } } }, { "id": "ai-social-app", "name": "AI Social App", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "ai-for-consumers" }, { "to": "chatbot" }, { "to": "digital-human" }, { "to": "ai-companion-risk" } ], "track": "today", "i18n": { "en": { "fullName": "AI Social App", "factExplain": "An app with AI characters inside social chats.", "humanExplain": "An AI social app is like a new kid in the group chat. It laughs at every joke, but its heart may be just Wi‑Fi.\n\nIt powers chat buddies and virtual characters. It also joins community hangouts. Set time limits. Keep private stuff private.", "humanExplainDisplay": "An AI social app is like\n==a new kid in the group chat==.\nIt laughs at every joke,\nbut its heart may be ==just Wi‑Fi==.\n\nIt powers chat buddies and virtual characters.\nIt also joins community hangouts.\nSet time limits.\nKeep private stuff private.", "relationsNarrative": "AI for Consumers\nAn AI social app is a common AI product for everyday people.\n\nChatbot\nA chatbot gives it its basic talk.\n\nDigital human\nA digital human turns the chat partner into a visible character.\n\nCompanion-risk\nThe more it feels like a friend, the more we must watch for dependence and emotional control.", "relations": { "ai-for-consumers": { "label": "is built for …", "note": "It is a common way everyday people first use AI." }, "chatbot": { "label": "talks through …", "note": "Chatbots provide the basic back-and-forth talk." }, "digital-human": { "label": "dresses up as …", "note": "Digital humans make the AI feel like a visible character." }, "ai-companion-risk": { "label": "can raise …", "note": "Close AI chats can create dependence." } } }, "zh": { "fullName": "AI 社交应用", "factExplain": "把 AI 角色接入社交互动的应用。", "humanExplain": "AI 社交应用像相亲角新来的嘴替：句句接得上，心里有没有你难说。\n\n用于陪聊、虚拟角色和社区互动，需防沉迷与隐私泄露。", "humanExplainDisplay": "AI 社交应用像相亲角新来的嘴替：\n==句句接得上==，\n心里有没有你，\n==难说==。\n\n用于陪聊、虚拟角色，\n和社区互动，\n需防沉迷与隐私泄露。", "relationsNarrative": "AI For Consumers\nAI 社交应用是面向大众的典型 AI 消费产品。\n\nChatbot\n聊天机器人提供它最基础的对话能力。\n\nDigital Human\n数字人把聊天对象包装成可见角色。\n\nCompanion Risk\n越像朋友，越要防依赖和情感操控。", "relations": { "ai-for-consumers": { "label": "面向…用户", "note": "它是大众 AI 的常见入口。" }, "chatbot": { "label": "用…对话", "note": "聊天机器人支撑核心互动。" }, "digital-human": { "label": "包装成…", "note": "数字人让 AI 更像角色。" }, "ai-companion-risk": { "label": "放大…风险", "note": "亲密陪聊可能带来依赖。" } } } } }, { "id": "ai-subscription-subsidies", "name": "AI Subsidies", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "ai-usage-gap" }, { "to": "ai-for-consumers" }, { "to": "ai-monetization" }, { "to": "ai-adoption-curve" } ], "track": "today", "i18n": { "en": { "fullName": "AI Subscription Subsidies", "factExplain": "A discount program that helps pay for an AI subscription.", "humanExplain": "An AI subscription subsidy is like a coupon for the fancy snack bar. Full price says no. Half off says yes to curly fries.\n\nIt lowers the cost of trying paid AI tools. You see it in work benefits and school AI programs.", "humanExplainDisplay": "An AI subscription subsidy is like\n==a coupon for the fancy snack bar==.\nFull price says no.\n==Half off== says yes to curly fries.\n\nIt lowers the cost\nof trying paid AI tools.\nYou see it in work benefits\nand school AI programs.", "relationsNarrative": "AI usage gap\nSubscription subsidies lower the price barrier, so more people may try AI.\n\nAI for Consumers\nSubsidies make it easier for people to use paid AI in daily life.\n\nAI monetization\nSubsidies change how platforms win users, keep users, and set prices.\n\nAdoption Curve\nLower prices can help AI spread sooner.", "relations": { "ai-usage-gap": { "label": "narrows …", "note": "Subsidies lower the price, but they do not teach AI skills." }, "ai-for-consumers": { "label": "helps spread …", "note": "Cheaper personal plans make daily AI trials easier." }, "ai-monetization": { "label": "shapes …", "note": "Subsidies can change sign-ups, renewals, and prices." }, "ai-adoption-curve": { "label": "speeds up …", "note": "Lower prices help early users spread the tool faster." } } }, "zh": { "fullName": "AI 订阅补贴", "factExplain": "为用户分担 AI 订阅费用的推广措施。", "humanExplain": "AI 订阅补贴像公司培训报销：原价肉疼不报名，能报销就敢去听课。\n\n降低个人和企业试用成本，常用于福利、教育、普及项目。", "humanExplainDisplay": "AI 订阅补贴像\n==公司培训报销==：\n原价肉疼不报名，\n能报销就敢去听课。\n\n降低个人和企业试用成本，\n常用于福利、教育、\n普及项目。", "relationsNarrative": "AI Usage Gap\n订阅补贴降低价格门槛，可能缩小使用差距。\n\nAI For Consumers\n补贴让个人更容易把 AI 订阅纳入日常。\n\nAI Monetization\n补贴会改变平台拉新、续费和定价策略。\n\nAdoption Curve\n价格门槛下降，会推动更早普及。", "relations": { "ai-usage-gap": { "label": "缩小…", "note": "补贴降门槛，但不会自动教会用户。" }, "ai-for-consumers": { "label": "推动…普及", "note": "个人订阅越便宜，日常尝试越容易。" }, "ai-monetization": { "label": "改变…策略", "note": "补贴会影响拉新、留存和定价。" }, "ai-adoption-curve": { "label": "加速…", "note": "价格下降能推动早期用户扩散。" } } } } }, { "id": "ai-super-app", "name": "AI super app", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "personal-ai-apps" }, { "to": "kimi-work" }, { "to": "api" }, { "to": "agent" } ], "track": "today", "i18n": { "en": { "fullName": "AI super app", "factExplain": "An app that puts many AI tools behind one main entrance.", "humanExplain": "An AI super app wants to be one big front door for AI. You stop hopping between apps like a frog with Wi-Fi.\n\nYou meet it in office tools and creator apps. AI helpers use it to handle many jobs from one place.", "humanExplainDisplay": "An AI super app wants to be\n==one big front door== for AI.\nYou stop hopping between apps\nlike a ==frog with Wi-Fi==.\n\nYou meet it in office tools\nand creator apps.\nAI helpers use it to handle\nmany jobs from one place.", "relationsNarrative": "Personal AI apps\nAn AI super app puts many Personal AI apps in one place.\n\nKimi Work\nKimi Work is a common work-style form of an AI super app.\n\nAPI\nAn AI super app often uses APIs to connect different models.\n\nAgent\nWith an Agent inside, it feels less like a toolbox and more like a helper.", "relations": { "personal-ai-apps": { "label": "bundles …", "note": "It puts many personal AI tools in one place." }, "kimi-work": { "label": "often takes … shape", "note": "Office products are a common way to build it." }, "api": { "label": "connects models through …", "note": "APIs often connect its many back-end models." }, "agent": { "label": "can include …", "note": "Agent skills make it feel like a helper that acts." } } }, "zh": { "fullName": "AI 超级应用", "factExplain": "把多种 AI 功能整合进同一入口的应用形态。", "humanExplain": "它想当 AI 界的商场中庭：问路、吃饭、买衣服都别出楼，一进门就把你的高频需求全包圆。\n\n常见于办公、创作和助手场景，目标是一个入口承接多种 AI 任务。", "humanExplainDisplay": "它想当 AI 界的==商场中庭==：\n问路、吃饭、买衣服都别出楼，\n一进门就把你的\n==高频需求全包圆==。\n\n常见于办公、创作\n和助手场景，\n目标是一个入口承接\n多种 AI 任务。", "relationsNarrative": "Personal AI apps\n它常把多个个人 AI 功能整合到同一个入口里。\n\nKimi Work\n办公型产品是 AI 超级应用的一种典型落地形态。\n\nAPI\n它表面是一个应用，底层常靠 API 接入不同模型。\n\nAgent\n加入代理能力后，它会从工具箱更像会做事的助手。", "relations": { "personal-ai-apps": { "label": "整合…能力", "note": "它常把多个个人 AI 功能装进一处。" }, "kimi-work": { "label": "常做成…形态", "note": "办公入口型产品是常见落地方式。" }, "api": { "label": "靠…接模型", "note": "前台像一个应用，后台常接多种模型。" }, "agent": { "label": "可把…塞进去", "note": "若能主动执行任务，体验会更像助手。" } } } } }, { "id": "ai-swipe-typing", "name": "AI Swipe Typing", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "language-modeling" }, { "to": "personal-ai-apps" } ], "track": "today", "i18n": { "en": { "fullName": "AI Swipe Typing", "factExplain": "AI predicts words from your swipe path and the sentence around it.", "humanExplain": "AI Swipe Typing is like a phone keyboard playing connect-the-dots. Your thumb draws a wobbly noodle, and it guesses the word.\n\nYou meet it in phone keyboards. It helps with one-hand typing and fixes words by using context.", "humanExplainDisplay": "AI Swipe Typing is like\na phone keyboard playing ==connect-the-dots==.\nYour thumb draws a ==wobbly noodle==,\nand it guesses the word.\n\nYou meet it in phone keyboards.\nIt helps with one-hand typing\nand fixes words by using context.", "relationsNarrative": "NLP\nAI Swipe Typing uses NLP to turn finger paths into word choices.\n\nLM\nLM helps it guess words from context and fix mistakes.\n\nPersonal AI apps\nAI Swipe Typing is a common input door for Personal AI apps.", "relations": { "natural-language-processing": { "label": "uses … to read text", "note": "NLP turns the swipe path into word choices." }, "language-modeling": { "label": "uses … to guess next words", "note": "LM uses context to make better word choices." }, "personal-ai-apps": { "label": "serves as an input door for …", "note": "The keyboard is a daily doorway into personal AI." } } }, "zh": { "fullName": "AI 滑行输入", "factExplain": "根据滑动轨迹和上下文预测输入文字。", "humanExplain": "AI 滑行输入像煎饼摊熟手：你手指绕两圈，它就摊出一串字。\n\n常见于手机输入法，适合单手输入，也会按上下文纠错。", "humanExplainDisplay": "AI 滑行输入像\n==煎饼摊熟手==：\n你手指绕两圈，\n它就==摊出一串字==。\n\n常见于手机输入法，\n适合单手输入，\n也会按上下文纠错。", "relationsNarrative": "NLP\n它用 NLP 把手势轨迹转成候选文字。\n\nLanguage Modeling\n语言模型提供上下文预测和纠错能力。\n\nPersonal AI Apps\n它是个人 AI 应用里的高频输入入口。", "relations": { "natural-language-processing": { "label": "依赖…处理文本", "note": "把滑动轨迹转成候选词。" }, "language-modeling": { "label": "用…预测下文", "note": "上下文越准，候选词越稳。" }, "personal-ai-apps": { "label": "作为…输入入口", "note": "键盘是贴身的 AI 入口。" } } } } }, { "id": "ai-talent-war", "name": "AI Talent War", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "compute-race" }, { "to": "frontier-model" }, { "to": "ai-anxiety" }, { "to": "ai-career-moat" } ], "track": "today", "i18n": { "en": { "fullName": "AI Talent War", "factExplain": "A fierce hiring fight for people with key AI skills.", "humanExplain": "It is like every school team chasing the same star pitcher. His cleats are still untied, and the pizza bribes are already out.\n\nIt pushes up pay and poaching in big model labs. Chip teams and research groups feel it too.", "humanExplainDisplay": "It is like ==every school team==\nchasing the ==same star pitcher==.\nHis cleats are still untied,\nand the pizza bribes are already out.\n\nIt pushes up pay and poaching\nin big model labs.\nChip teams and research groups\nfeel it too.", "relationsNarrative": "Compute-race\nThe talent fight and the compute fight often grow together.\n\nFrontier model\nFrontier model work needs top researchers, so the hiring fight gets fierce.\n\nAI-anxiety\nHuge pay stories can make normal workers feel behind.\n\nAI career moat\nA fierce talent war makes a strong career moat more useful.", "relations": { "compute-race": { "label": "pushes up … too", "note": "Talent fights and compute fights often heat up together." }, "frontier-model": { "label": "centers on …", "note": "Frontier model teams depend on top researchers most." }, "ai-anxiety": { "label": "fuels …", "note": "Big pay stories can make normal workers worry." }, "ai-career-moat": { "label": "pushes people toward …", "note": "A hotter talent war makes hard-to-replace skills matter more." } } }, "zh": { "fullName": "AI 人才争夺战", "factExplain": "围绕 AI 关键人才展开的激烈招聘竞争。", "humanExplain": "像武侠片里各大门派抢关门弟子：人还没出山，offer 和签字费先飞满天。\n\n它会推高薪资和挖人力度，常见于大模型、芯片和研究团队。", "humanExplainDisplay": "像武侠片里\n各大门派抢\n==关门弟子==：\n人还没出山，offer\n和签字费先\n==飞满天==。\n\n它会推高薪资\n和挖人力度，\n常见于大模型、芯片\n和研究团队。", "relationsNarrative": "Compute-race\n抢人才和抢算力，常常是同一场竞赛的两面。\n\nFrontier model\n前沿模型研发最吃顶尖人才，争夺通常最激烈。\n\nAI-anxiety\n高薪挖人新闻容易放大普通人的职业焦虑。\n\nAI career moat\n人才战越凶，个人越需要建立难替代的护城河。", "relations": { "compute-race": { "label": "和…互相推高", "note": "抢人才与抢算力常同步升级。" }, "frontier-model": { "label": "围绕…展开", "note": "前沿模型公司最依赖顶尖研究者。" }, "ai-anxiety": { "label": "加剧…情绪", "note": "普通人易被高薪神话带焦虑。" }, "ai-career-moat": { "label": "倒逼建立…", "note": "竞争越激烈，越要做难替代能力。" } } } } }, { "id": "ai-toggle", "name": "AI Toggle", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "ai-for-consumers" }, { "to": "permission-fatigue" }, { "to": "ai-file-permissions" } ], "track": "today", "i18n": { "en": { "fullName": "AI Toggle", "factExplain": "A product switch that lets users turn AI features on or off.", "humanExplain": "An AI Toggle is like a light switch for a robot helper. Turn it on for help, and off before it names your houseplants.\n\nYou see it in work apps. You see it in search and photo apps. It lets you decide when AI joins in.", "humanExplainDisplay": "An AI Toggle is like a ==light switch==\nfor a ==robot helper==.\nTurn it on for help,\nand off before it names your houseplants.\n\nYou see it in work apps.\nYou see it in search and photo apps.\nIt lets you decide when AI joins in.", "relationsNarrative": "AI for Consumers\nAn AI Toggle adds optional control to everyday AI products.\n\nPermission fatigue\nToo many AI Toggles can make choice feel like a bother.\n\nAI File Permissions\nAI File Permissions set what AI can reach after the toggle is on.", "relations": { "ai-for-consumers": { "label": "adds choice to …", "note": "AI Toggles make everyday products feel more controllable." }, "permission-fatigue": { "label": "can worsen …", "note": "Too many switches can turn control into a chore." }, "ai-file-permissions": { "label": "works with …", "note": "File permissions set what AI can reach after the switch is on." } } }, "zh": { "fullName": "AI 开关", "factExplain": "让用户启用或关闭 AI 功能的产品开关。", "humanExplain": "AI 开关像小区门禁遥控：想帮忙就开门，推销员一来立刻关门。\n\n用于办公、搜索、相册，让人按场景决定 AI 是否上场。", "humanExplainDisplay": "AI 开关像==小区门禁遥控==：\n想帮忙就开门，\n推销员一来\n==立刻关门==。\n\n用于办公、搜索、相册，\n让人按场景决定\nAI 是否上场。", "relationsNarrative": "AI For Consumers\nAI 开关给大众产品加一层可选控制。\n\nPermission Fatigue\n开关太多时，选择权也会变成打扰。\n\nAI File Permissions\n文件权限限定 AI 开关打开后的可触达范围。", "relations": { "ai-for-consumers": { "label": "给…加选择权", "note": "AI 开关让大众产品更可控。" }, "permission-fatigue": { "label": "可能加重…", "note": "开关太多会把控制感变成负担。" }, "ai-file-permissions": { "label": "配合…控范围", "note": "权限决定 AI 能碰哪些文件。" } } } } }, { "id": "ai-trade-secrets", "name": "AI Trade Secrets", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "closed-model" }, { "to": "data-exfiltration" }, { "to": "anti-distillation" }, { "to": "copyright" } ], "track": "today", "seo": { "zh": { "title": "AI 商业秘密 是什么?锁后厨的底料,一文看懂 — AI Rookies", "description": "企业不公开的模型、数据和训练方法机密。人话版一分钟讲明白。" }, "en": { "title": "What Is AI Trade Secrets? Secret Sauce in a Locked Drawer", "description": "Private AI model, data, and training secrets a company keeps hidden. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "AI Trade Secrets", "factExplain": "Private AI model, data, and training secrets a company keeps hidden.", "humanExplain": "AI trade secrets are the burger place's secret sauce. The line wraps around the block, but the recipe sleeps in a locked drawer.\n\nThey guide what gets shared. They matter in secret deals and lawsuits.", "humanExplainDisplay": "AI trade secrets are the ==burger place's secret sauce==.\nThe line wraps around the block,\nbut the ==recipe sleeps in a locked drawer==.\n\nThey guide what gets shared.\nThey matter in secret deals and lawsuits.", "relationsNarrative": "Closed-source Model\nTrade secrets often explain why closed models hide weights and training details.\n\nExfiltration\nA leak can carry away model recipes and customer data.\n\nAnti-distillation\nAnti-distillation tries to stop others from copying model skills through the API.\n\nCopyright\nCopyright protects works, but trade secrets protect hidden information.", "relations": { "closed-model": { "label": "supports …", "note": "Trade secrets help closed models hide weights and training details." }, "data-exfiltration": { "label": "guards against …", "note": "A leak can carry away model recipes and customer data." }, "anti-distillation": { "label": "leads to …", "note": "Anti-distillation tries to stop others from copying model skills through the API." }, "copyright": { "label": "differs from …", "note": "Copyright protects works. Trade secrets protect hidden information." } } }, "zh": { "fullName": "AI 商业秘密", "factExplain": "企业不公开的模型、数据和训练方法机密。", "humanExplain": "AI 商业秘密就是火锅店底料包：店能开满城，配方锁后厨，谁偷走就能仿味。\n\n它决定开源边界，也牵动保密、合作审查和诉讼。", "humanExplainDisplay": "AI 商业秘密就是\n==火锅店底料包==：\n店能开满城，\n配方锁后厨，谁偷走就能仿味。\n\n它决定开源边界，\n也牵动保密、合作审查\n和诉讼。", "relationsNarrative": "Closed-source Model\n商业秘密常是闭源模型不公开权重和训练细节的理由。\n\nExfiltration\n一旦外泄，模型配方、客户数据都可能被带走。\n\nAnti-distillation\n反蒸馏试图防别人从接口反推出模型能力。\n\nCopyright\n版权保护表达，商业秘密保护未公开信息。", "relations": { "closed-model": { "label": "支撑…", "note": "商业秘密常是闭源的重要护城河。" }, "data-exfiltration": { "label": "防范…", "note": "外泄会带走模型配方和客户资料。" }, "anti-distillation": { "label": "催生…", "note": "反蒸馏防别人套走模型能力。" }, "copyright": { "label": "区别于…", "note": "版权护作品，商业秘密护未公开信息。" } } } } }, { "id": "ai-travel-agent", "name": "AI Travel Agent", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "agentic-commerce" }, { "to": "function-call" } ], "track": "today", "i18n": { "en": { "fullName": "AI Travel Agent", "factExplain": "An AI app that can plan, book, and adjust your trip.", "humanExplain": "An AI Travel Agent is the friend who makes the group trip happen. You give dates and a budget. It herds flights and hotels like cats.\n\nIt books flights and hotels, checks prices, and changes plans. Keep a close eye on payment approval.", "humanExplainDisplay": "An AI Travel Agent is ==the friend who makes the group trip happen==.\nYou give dates and a budget.\nIt ==herds flights and hotels like cats==.\n\nIt books flights and hotels,\nchecks prices,\nand changes plans.\nKeep a close eye on payment approval.", "relationsNarrative": "Agent\nAn AI Travel Agent is an Agent used for travel planning.\n\nAgentic commerce\nFlight and hotel booking puts it in Agentic commerce.\n\nFunction-calling\nIt uses Function-calling to reach flight, hotel, and map tools.", "relations": { "agent": { "label": "uses … for travel", "note": "It turns Agent planning and action into a travel plan." }, "agentic-commerce": { "label": "books through …", "note": "Booking flights and hotels means ordering for the user." }, "function-call": { "label": "uses … to call tools", "note": "It calls flight, hotel, and map tools." } } }, "zh": { "fullName": "AI 旅行代理", "factExplain": "能规划、预订并调整旅行行程的 AI 应用。", "humanExplain": "AI 旅行代理像懂省钱的驴友队长：你报钱包和假期，它把机酒路线拉群排好。\n\n用于订票订房、改行程和比价，付款授权要盯紧。", "humanExplainDisplay": "AI 旅行代理像\n懂省钱的==驴友队长==：\n你报钱包和假期，\n它把机酒路线==拉群排好==。\n\n用于订票订房、改行程和比价，\n付款授权要盯紧。", "relationsNarrative": "Agent\n它是 Agent 在旅行规划里的落地形态。\n\nAgentic commerce\n订票订房让它进入替用户下单的商业场景。\n\nFunction-calling\n它靠 Function-calling 调航班、酒店、地图接口。", "relations": { "agent": { "label": "把…用于旅行", "note": "把规划和执行落到行程上。" }, "agentic-commerce": { "label": "接入…下单", "note": "订票订房本质是替人下单。" }, "function-call": { "label": "用…调接口", "note": "调用航班、酒店和地图接口。" } } } } }, { "id": "ai-tutor", "name": "AI Tutor", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "prompt" }, { "to": "rag" }, { "to": "hallucination" } ], "track": "today", "i18n": { "en": { "fullName": "AI Tutor", "factExplain": "An AI app that gives personal lessons, practice, and study help.", "humanExplain": "An AI Tutor is the homework buddy who never says, “Figure it out yourself.” It will explain fractions again while your real friend raids the fridge.\n\nYou meet it in study apps and chatbots for lessons and practice. It can plan your study time too, but it may sound sure and still be wrong.", "humanExplainDisplay": "An AI Tutor is the ==homework buddy==\nwho never says,\n“Figure it out yourself.”\nIt will ==explain fractions again==\nwhile your real friend raids the fridge.\n\nYou meet it in study apps and chatbots\nfor lessons and practice.\nIt can plan your study time too,\nbut it may sound sure\nand still be wrong.", "relationsNarrative": "LLM\nAn AI Tutor usually uses an LLM to understand questions and write explanations.\n\nPrompt\nA Prompt sets its teaching style, difficulty, and answer format.\n\nRAG\nRAG can connect it to textbooks and question banks for a course.\n\nHallucination\nHallucination means it can sound confident and still teach the wrong answer.", "relations": { "llm": { "label": "usually runs on …", "note": "Most AI Tutors use an LLM to talk and explain." }, "prompt": { "label": "gets its teaching style from …", "note": "The Prompt sets how it teaches and how detailed it gets." }, "rag": { "label": "can use … for course facts", "note": "RAG keeps its answers closer to the class material." }, "hallucination": { "label": "can fall into …", "note": "It may sound confident and still explain a problem wrong." } } }, "zh": { "fullName": "AI 导师", "factExplain": "一种提供个性化讲解与陪练的教学型 AI 应用。", "humanExplain": "AI 导师就像宿舍里那个永不熄灯的学霸室友：你哪道题卡壳，它就能陪你反复掰到明白。\n\n它常拿来讲题、陪练和定计划，但也可能看着自信却讲错。", "humanExplainDisplay": "AI 导师就像宿舍里\n那个==永不熄灯的学霸室友==：\n你哪道题卡壳，\n它就能陪你\n==反复掰到明白==。\n\n它常拿来讲题、\n陪练和定计划，\n但也可能看着自信却讲错。", "relationsNarrative": "LLM\n它通常建立在大模型之上，负责理解问题并生成讲解。\n\nPrompt\n提示词会设定它的教学风格、难度和回答方式。\n\nRAG\n接入教材、题库后，它的回答会更贴近特定课程内容。\n\nHallucination\n它也会自信讲错，所以不该把它当绝对标准答案。", "relations": { "llm": { "label": "通常基于…", "note": "大多数这类产品由大模型驱动对话。" }, "prompt": { "label": "靠…定人设", "note": "提示词决定它怎么教、教到多细。" }, "rag": { "label": "可接入…补课", "note": "接教材题库后，讲解会更贴课纲。" }, "hallucination": { "label": "容易出现…", "note": "讲题看着自信，未必真的讲对。" } } } } }, { "id": "ai-unit-economics-ai", "name": "AI Unit Economics", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "cost-aware-ai-ai" }, { "to": "tokens-per-second" }, { "to": "continuous-batching" }, { "to": "ai-monetization" } ], "track": "today", "i18n": { "en": { "fullName": "AI Unit Economics", "factExplain": "A way to see if one AI use costs less than it earns.", "humanExplain": "Picture a lemonade stand with a line around the block. If each cup sells for $1 but costs $1.25, success is just a faster way to go broke.\n\nAI unit economics checks the cost and income for one AI task. Teams use it for support bots and coding helpers before they grow.", "humanExplainDisplay": "Picture a ==lemonade stand==\nwith a line around the block.\nIf each cup sells for $1\nbut costs $1.25,\nsuccess is just a ==faster way to go broke==.\n\nAI unit economics checks\nthe cost and income\nfor one AI task.\nTeams use it for support bots\nand coding helpers\nbefore they grow.", "relationsNarrative": "Cost-aware AI\nUnit economics shows Cost-aware AI where to spend less.\n\nTPS\nHigher TPS can handle more requests in the same time.\n\nContinuous batching\nContinuous batching cuts idle time, so one request can cost less.\n\nAI monetization\nAI monetization sets the price, but unit economics says if it still earns money.", "relations": { "cost-aware-ai-ai": { "label": "guides … choices", "note": "Clear unit costs show the system where to spend less." }, "tokens-per-second": { "label": "is affected by …", "note": "Higher TPS can lower the cost of each request." }, "continuous-batching": { "label": "cuts cost with …", "note": "It reduces idle compute, so each request can cost less." }, "ai-monetization": { "label": "decides whether … works", "note": "A good price model still needs each request to make money." } } }, "zh": { "fullName": "AI 单位经济模型", "factExplain": "衡量单次 AI 服务成本与收益是否划算的方法。", "humanExplain": "算这个，像食堂承包窗口：打饭的人再挤，每份都倒贴两块，阿姨手抡出残影，也只是越忙越亏。\n\n常用来评估客服、代码助手等 AI 服务，判断能否规模化赚钱。", "humanExplainDisplay": "算这个，像==食堂承包窗口==：\n打饭的人再挤，\n每份都倒贴两块，\n也只是==越忙越亏==。\n\n常用来评估客服、\n代码助手等 AI 服务，\n判断能否\n规模化赚钱。", "relationsNarrative": "Cost-aware AI\n先把单次账算明白，系统优化才知道该省哪儿。\n\nTokens per second\n生成速度越高，单位时间能服务更多请求。\n\nContinuous batching\n它通过减少空转，直接改善单次服务成本。\n\nAI monetization\n变现讲的是怎么收费，它讲的是收了还赚不赚。", "relations": { "cost-aware-ai-ai": { "label": "指导…取舍", "note": "先算清成本，系统才会学会省着用。" }, "tokens-per-second": { "label": "受…影响", "note": "速度更高，往往能摊薄单次服务成本。" }, "continuous-batching": { "label": "靠…降本", "note": "减少算力空转，常直接改善毛利。" }, "ai-monetization": { "label": "决定…能否成立", "note": "变现模式再好，也得先跑通单次账。" } } } } }, { "id": "ai-usage-cap", "name": "Usage cap", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "inference" }, { "to": "gpu" }, { "to": "tokens-per-second" }, { "to": "ai-data-center" } ], "track": "today", "i18n": { "en": { "fullName": "AI usage cap", "factExplain": "A limit on how much AI you can use in a set time.", "humanExplain": "A usage cap is like a pizza shop on Friday night. You want a third slice, but the oven says, “Buddy, I need a nap.”\n\nYou see it as chat limits or time with the best model. It saves money and keeps the service from jamming.", "humanExplainDisplay": "A usage cap is like a ==pizza shop== on Friday night.\nYou want a third slice,\nbut the oven says,\n==“Buddy, I need a nap.”==\n\nYou see it as chat limits\nor time with the best model.\nIt saves money\nand keeps the service from jamming.", "relationsNarrative": "Inference\nA usage cap limits how much inference power users can take.\n\nGPU\nWhen GPU supply is tight, platforms are more likely to add caps.\n\nTPS\nBeyond the quota, generation speed also changes how the service feels.\n\nAI data center\nA usage cap often reflects the cost pressure behind the data center.", "relations": { "inference": { "label": "limits … use", "note": "Each AI answer uses inference resources." }, "gpu": { "label": "depends on … supply", "note": "Tight GPU supply makes caps more likely." }, "tokens-per-second": { "label": "trades off with …", "note": "Speed and quota both shape how the AI feels." }, "ai-data-center": { "label": "reflects … cost", "note": "Caps often show the pressure of data center costs." } } }, "zh": { "fullName": "AI 使用上限", "factExplain": "平台对用户在一定时间内的 AI 使用量限制。", "humanExplain": "跟热门奶茶店发号一个道理：不是你不想点了，是系统先说“今天先排到这儿”。\n\n常限制聊天次数或高阶模型额度，用来控成本并防服务拥堵。", "humanExplainDisplay": "跟热门奶茶店==发号==一个道理：\n不是你不想点了，\n是系统先说==“今天先排到这儿”==。\n\n常限制聊天次数或高阶模型额度，\n用来控成本并防服务拥堵。", "relationsNarrative": "Inference\n使用上限本质是在限制推理资源的占用。\n\nGPU\nGPU 供给越紧，平台越可能提高限制。\n\nTokens-per-second\n额度之外，生成速度也会影响用户体感。\n\nAI data center\n使用上限常映射数据中心的成本压力。", "relations": { "inference": { "label": "限制…消耗", "note": "每次生成都会占用推理资源。" }, "gpu": { "label": "受…供给影响", "note": "算力紧张时更容易设上限。" }, "tokens-per-second": { "label": "常与…权衡", "note": "速度和额度常一起影响体验。" }, "ai-data-center": { "label": "反映…成本", "note": "背后是机房与算力投入压力。" } } } } }, { "id": "ai-usage-gap", "name": "AI usage gap", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "ai-anxiety" }, { "to": "ai-adoption-curve" }, { "to": "ai-career-moat" } ], "track": "today", "i18n": { "en": { "fullName": "AI usage gap", "factExplain": "The speed gap between skilled AI users and people still doing it by hand.", "humanExplain": "It is a homework race with one kid on a smart tutor app. You are still hunting for a pencil.\n\nIt makes work and study speeds drift apart. You see it at work. You see it in writing. You see it when looking things up.", "humanExplainDisplay": "It is a ==homework race==\nwith one kid on a smart tutor app.\nYou are still ==hunting for a pencil==.\n\nIt makes work and study speeds drift apart.\nYou see it at work.\nYou see it in writing.\nYou see it when looking things up.", "relationsNarrative": "AI-anxiety\nSeeing others work faster with AI can make you feel behind.\n\nAdoption Curve\nThe gap is how the Adoption Curve feels in real life.\n\nAI career moat\nPeople who use AI better can build a stronger edge.", "relations": { "ai-anxiety": { "label": "can worsen …", "note": "Seeing others use AI smoothly can make you doubt yourself." }, "ai-adoption-curve": { "label": "shows the split in …", "note": "Early users often get the speed boost first." }, "ai-career-moat": { "label": "pushes people to build …", "note": "Knowing how to use AI is becoming a new edge." } } }, "zh": { "fullName": "AI 使用鸿沟", "factExplain": "会用 AI 与不会用 AI 之间的效率差距。", "humanExplain": "这差距像考试时有人带了会押题的学霸笔记，你还在翻书找重点；铃一响，分数和心态一起被甩开。\n\n它会放大工作学习中的效率差，常见于办公、写作和查资料。", "humanExplainDisplay": "这差距像考试时\n有人带了==会押题的学霸笔记==，\n你还在翻书找重点；\n铃一响，==分数和心态一起被甩开==。\n\n它会放大工作学习中的\n效率差，\n常见于办公、写作\n和查资料。", "relationsNarrative": "AI-anxiety\n看见别人用 AI 提效，常会放大自己的落后感。\n\nAdoption Curve\n它是采用曲线拉开后的真实个人体感。\n\nAI career moat\n谁更会用 AI，谁就更容易形成新优势。", "relations": { "ai-anxiety": { "label": "会放大…", "note": "越看别人用得顺，越容易自我怀疑。" }, "ai-adoption-curve": { "label": "体现…分化", "note": "早用熟的人，往往先吃到效率红利。" }, "ai-career-moat": { "label": "倒逼建立…", "note": "会用 AI 的方法，正变成新优势。" } } } } }, { "id": "ai-vendor-lock-in", "name": "AI Vendor Lock-in", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "api" }, { "to": "closed-model" }, { "to": "open-source-model" }, { "to": "on-premise-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI Vendor Lock-in", "factExplain": "The risk of being stuck with one AI vendor because moving is hard.", "humanExplain": "AI vendor lock-in is like renting a moving truck with your sofa bolted inside. Want a new truck? First pay to unbolt the sofa.\n\nYou meet it with cloud AI models and vendor-only APIs. It can raise costs and make migration painful.", "humanExplainDisplay": "AI vendor lock-in is like renting ==a moving truck==\nwith your ==sofa bolted inside==.\nWant a new truck?\nFirst pay to unbolt the sofa.\n\nYou meet it with cloud AI models\nand vendor-only APIs.\nIt can raise costs\nand make migration painful.", "relationsNarrative": "API\nA vendor-only API is often the front door to vendor lock-in.\n\nClosed-source Model\nClosed-source models hide details, so switching can cost more.\n\nOpen-source-model\nOpen-source models give teams more choices and more bargaining power.\n\nOn-premise AI\nOn-premise AI keeps data and systems under your control.", "relations": { "api": { "label": "locks you in through …", "note": "The deeper you use a vendor-only API, the harder moving gets." }, "closed-model": { "label": "is worsened by …", "note": "Closed-source models make replacement more costly." }, "open-source-model": { "label": "can be reduced with …", "note": "Open-source models leave you more ways to switch." }, "on-premise-ai": { "label": "keeps control with …", "note": "On-premise AI reduces ties to outside vendors." } } }, "zh": { "fullName": "AI 供应商锁定", "factExplain": "因依赖特定厂商而难以迁移的风险。", "humanExplain": "AI 供应商锁定像办健身房年卡：柜钥匙在人家手里，想换馆还得赎装备。\n\n常见于云模型和专有 API，影响成本与迁移。", "humanExplainDisplay": "AI 供应商锁定像办\n==健身房年卡==：\n柜钥匙在人家手里，\n想换馆还得==赎装备==。\n\n常见于云模型和专有 API，\n影响成本与迁移。", "relationsNarrative": "API\n专有 API 往往是供应商锁定最直接的入口。\n\nClosed-source Model\n闭源模型不开放细节，会加重替换和迁移成本。\n\nOpen-source-model\n开源模型能给企业更多替代方案和议价空间。\n\nOn-premise AI\n本地部署让数据和运行环境更少受厂商牵制。", "relations": { "api": { "label": "通过…绑定入口", "note": "专有 API 越深，迁移越费劲。" }, "closed-model": { "label": "常由…加重", "note": "闭源模型让替换成本更高。" }, "open-source-model": { "label": "可用…降低依赖", "note": "开源模型给迁移留后路。" }, "on-premise-ai": { "label": "借…保留控制权", "note": "本地部署能减少外部绑定。" } } } } }, { "id": "ai-video-indexing", "name": "Video-indexing", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "multimodal" }, { "to": "vector-search" }, { "to": "ocr" } ], "track": "today", "i18n": { "en": { "fullName": "AI Video Indexing", "factExplain": "A way for AI to tag video content so you can search it.", "humanExplain": "AI video indexing is like a friend with sticky notes watching a birthday video. It marks “cake drop at 4:03” and “dog steals hat at 5:10.”\n\nIt helps you find the right video moment fast. You meet it in security review, media archives, and company search.", "humanExplainDisplay": "AI video indexing is like\na friend with ==sticky notes==\nwatching a birthday video.\nIt marks ==“cake drop at 4:03”==\nand “dog steals hat at 5:10.”\n\nIt helps you find\nthe right video moment fast.\nYou meet it in security review,\nmedia archives,\nand company search.", "relationsNarrative": "Computer Vision\nAI video indexing uses Computer Vision to understand video scenes.\n\nMultimodal AI\nAI video indexing often uses pictures, speech, and text together.\n\nVector search\nVector search helps it find related clips by meaning.\n\nOCR\nOCR reads subtitles and signs in the video for indexing.", "relations": { "computer-vision": { "label": "uses …", "note": "It uses Computer Vision to understand what appears on screen." }, "multimodal": { "label": "combines … signals", "note": "It often indexes pictures, speech, and text together." }, "vector-search": { "label": "finds clips with …", "note": "Vector search helps find video clips by meaning, not exact words." }, "ocr": { "label": "reads on-screen text with …", "note": "OCR can pull subtitles, signs, and labels from the video." } } }, "zh": { "fullName": "AI Video Indexing", "factExplain": "用 AI 自动识别并标注视频内容的检索方法。", "humanExplain": "像给食堂后厨装了个会记账的老板娘：哪锅几点出炉、谁端走了什么、辣不辣，全给你记得明明白白。\n\n能帮你快速定位视频片段，常用于安防排查、媒体归档和企业检索。", "humanExplainDisplay": "像给食堂后厨装了个\n==会记账的老板娘==：\n哪锅几点出炉、谁端走了什么、\n辣不辣，全给你记得明明白白。\n\n能帮你快速定位视频片段，\n常用于安防排查、媒体归档\n和企业检索。", "relationsNarrative": "Computer Vision\n它是计算机视觉在视频理解与检索里的常见应用。\n\nMultimodal\n视频索引常同时利用画面、语音和文字信息。\n\nVector Search\n向量搜索让它能按语义查到相关视频片段。\n\nOCR\nOCR 可提取片中字幕和招牌等文字用于索引。", "relations": { "computer-vision": { "label": "属于…应用", "note": "它靠视觉模型理解画面内容。" }, "multimodal": { "label": "结合…信息", "note": "常把画面、语音、文字一起索引。" }, "vector-search": { "label": "用…找片段", "note": "把视频特征向量化后再检索。" }, "ocr": { "label": "提取…字幕字牌", "note": "画面里的字也常被纳入索引。" } } } } }, { "id": "ai-vulnerability-discovery-ai", "name": "AI Vulnerability Discovery", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "remote-code-execution" }, { "to": "agentic-coding" }, { "to": "alignment" }, { "to": "ai-biosecurity" } ], "track": "today", "i18n": { "en": { "fullName": "AI Vulnerability Discovery", "factExplain": "AI that automatically finds security holes in software or computer systems.", "humanExplain": "Vuln Discovery is like a nosy home inspector in your app. It jiggles every door and says, “Uh… this lock is decorative.”\n\nTeams use it in code reviews and security drills. It is faster, but it can miss holes or raise false alarms.", "humanExplainDisplay": "Vuln Discovery is like a ==nosy home inspector==\nin your app.\nIt ==jiggles every door== and says,\n“Uh… this lock is decorative.”\n\nTeams use it in code reviews\nand security drills.\nIt is faster,\nbut it can miss holes\nor raise false alarms.", "relationsNarrative": "RCE\nVuln Discovery often looks for RCE because it is a very dangerous bug.\n\nAgentic coding\nAgentic coding can write code and run tests, so it can speed up the hunt.\n\nAlignment\nAlignment helps keep this power aimed at defense, not abuse.\n\nAI biosecurity\nAI biosecurity also deals with dual-use risk from stronger AI.", "relations": { "remote-code-execution": { "label": "targets … risk", "note": "It often looks for high-risk RCE bugs." }, "agentic-coding": { "label": "is sped up by …", "note": "Agents that code and test can find holes faster." }, "alignment": { "label": "is limited by …", "note": "Stronger bug-finding power needs stronger misuse limits." }, "ai-biosecurity": { "label": "shares risk with …", "note": "Both are safety issues from dual-use AI power." } } }, "zh": { "fullName": "AI 漏洞发现", "factExplain": "用 AI 自动发现软件或系统安全漏洞。", "humanExplain": "这活儿像老中医把脉，不等你真倒下，先把系统里那些暗病、虚火和漏风口摸出来。\n\n常用于代码审计和攻防测试，排查更快，但仍会误报漏报。", "humanExplainDisplay": "这活儿像老中医==把脉==，\n不等你真倒下，\n先把系统里那些暗病、虚火\n和==漏风口==摸出来。\n\n常用于代码审计和攻防测试，\n排查更快，但仍会误报漏报。", "relationsNarrative": "Remote-code-execution\n它常被用来寻找远程执行这类高危漏洞。\n\nAgentic-coding\n会写代码也会跑测试的代理，能加速自动找洞。\n\nAlignment\n这类能力既能防守也能攻击，因此更依赖对齐约束。\n\nAI biosecurity\n两者都属于能力增强后带来的双用途安全风险。", "relations": { "remote-code-execution": { "label": "瞄准…风险", "note": "常用于发现高危远程执行漏洞。" }, "agentic-coding": { "label": "可被…加速", "note": "会写会跑的代理更适合自动找洞。" }, "alignment": { "label": "受…约束", "note": "能力越强，越需要限制滥用方向。" }, "ai-biosecurity": { "label": "同属高风险", "note": "都属于双用途能力带来的安全议题。" } } } } }, { "id": "ai-war-game-simulation", "name": "AI Wargaming Simulation", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "multi-agent-system" }, { "to": "world-model" }, { "to": "agent" }, { "to": "ai-biosecurity" } ], "track": "today", "i18n": { "en": { "fullName": "AI Wargaming Simulation", "factExplain": "AI wargaming uses AI to simulate rival teams and test strategies.", "humanExplain": "AI wargaming is a serious board game night. The AI plays every side, even the cousin who cheats at Monopoly.\n\nYou meet it in defense work. You also meet it in security and emergency drills. It helps teams find weak spots early.", "humanExplainDisplay": "AI wargaming is a ==serious board game night==.\nThe AI ==plays every side==,\neven the cousin who cheats at Monopoly.\n\nYou meet it in defense work.\nYou also meet it in security and emergency drills.\nIt helps teams find weak spots early.", "relationsNarrative": "MAS\nAI wargaming often uses a MAS to play the different sides.\n\nWorld model\nAI wargaming needs a world model to show how the situation changes.\n\nAgent\nEach role in the drill can be played by an Agent.\n\nAI biosecurity\nAI wargaming can test response plans for high-risk biosecurity problems.", "relations": { "multi-agent-system": { "label": "often runs as …", "note": "Many rival players are easier to simulate with a MAS." }, "world-model": { "label": "builds the scene with …", "note": "A world model lets the situation change over time." }, "agent": { "label": "puts … in the drill", "note": "An Agent can play one role and choose actions." }, "ai-biosecurity": { "label": "rehearses … risks", "note": "It can test response plans for risky bio scenarios." } } }, "zh": { "fullName": "AI 兵棋推演模拟", "factExplain": "用 AI 模拟多方博弈与策略对抗的系统。", "humanExplain": "还没真上场，先让 AI 把牌桌掀上千遍：谁会诈唬、谁先露怯，提前全试出来。\n\n常用于国防、安全和应急推演，帮助找策略漏洞。", "humanExplainDisplay": "还没真上场，\n先让 AI 把==牌桌掀上千遍==：\n谁会诈唬、谁先露怯，\n提前全==摸透路数==。\n\n常用于国防、\n安全和应急推演，\n帮助找策略漏洞。", "relationsNarrative": "Multi-Agent-System\n多方博弈场景里，它常用多智能体来扮演各方。\n\nWorld-Model\n推演要先有环境模型，才能模拟局势如何变化。\n\nAgent\n每个参演角色，常由能行动决策的代理承担。\n\nAI Biosecurity\n它可用于高风险议题的预演与应对策略测试。", "relations": { "multi-agent-system": { "label": "常做成…", "note": "多方对抗常靠多智能体来模拟。" }, "world-model": { "label": "依赖…建局面", "note": "需要可演化的环境与局势表征。" }, "agent": { "label": "把…放进推演", "note": "单个行动者常由代理来扮演。" }, "ai-biosecurity": { "label": "可用于预演…风险", "note": "能在高风险场景先做桌面推演。" } } } } }, { "id": "ai-website-cloning", "name": "AI Website Cloning", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "ai-app-builder" }, { "to": "spec-to-code" }, { "to": "ai-accelerated-prototyping" }, { "to": "copyright" } ], "track": "today", "i18n": { "en": { "fullName": "AI Website Cloning", "factExplain": "AI that makes a similar website from a page or screenshot.", "humanExplain": "AI Website Cloning is like copying a lunch tray. Same burger shape. Same fries. Please do not copy the ketchup stain too.\n\nPeople use it for mock sites and quick prototypes. Watch out for copied text, logos, and phishing tricks.", "humanExplainDisplay": "AI Website Cloning is like\n==copying a lunch tray==.\nSame burger shape.\nSame fries.\nPlease do not copy\n==the ketchup stain== too.\n\nPeople use it for mock sites\nand quick prototypes.\nWatch out for copied text,\nlogos, and phishing tricks.", "relationsNarrative": "AI App Builder\nWebsite cloning is often built into AI App Builders for a fast first draft.\n\nSpec-to-code\nIt treats a screenshot or page like a spec and makes draft front-end code.\n\nAI Prototyping\nIt builds a similar shell first, so people can test an idea fast.\n\nCopyright\nCopying the page look and words can cross a copyright line.", "relations": { "ai-app-builder": { "label": "plugs into …", "note": "AI App Builders use cloning to start a site fast." }, "spec-to-code": { "label": "turns into …", "note": "The cloned page still needs to become real front-end code." }, "ai-accelerated-prototyping": { "label": "speeds up …", "note": "Cloning the shell helps teams test an idea sooner." }, "copyright": { "label": "may cross …", "note": "Copying the look and words too closely can break copyright rules." } } }, "zh": { "fullName": "AI 网页克隆", "factExplain": "用 AI 从网页或截图生成相似网站。", "humanExplain": "AI 网页克隆像裁缝照旧衣打版：领口袖口学像，商标别也缝上，太露馅。\n\n用于仿站和原型，也要防版权、钓鱼风险。", "humanExplainDisplay": "AI 网页克隆像\n==裁缝照旧衣打版==：\n领口袖口学像，\n商标别也缝上，==太露馅==。\n\n用于仿站和原型，\n也要防版权、\n钓鱼风险。", "relationsNarrative": "AI App Builder\n网页克隆常被集成进建站工具，用来快速起稿。\n\nSpec-to-code\n它把截图或网页当规格，生成前端代码草稿。\n\nAI Prototyping\n它先搭出相似外壳，方便快速验证想法。\n\nCopyright\n照搬页面视觉和文案，容易踩到版权红线。", "relations": { "ai-app-builder": { "label": "嵌入…建站", "note": "建站工具常用它快速起稿。" }, "spec-to-code": { "label": "借…落代码", "note": "仿站结果最终要落成代码。" }, "ai-accelerated-prototyping": { "label": "加速…", "note": "先仿出外壳，再验证想法。" }, "copyright": { "label": "可能触碰…", "note": "照搬视觉和文案容易侵权。" } } } } }, { "id": "ai-winter", "name": "AI Winter", "layer": "L1", "era": "1974", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "dartmouth-workshop" }, { "to": "expert-system" }, { "to": "ai-bubble" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "AI Winter", "factExplain": "A time when AI loses money and attention after big promises fail.", "humanExplain": "AI Winter is the day after a school talent show flops. The fog machine gets packed away, and the pizza money disappears.\n\nIt explains AI’s cold spells after big promises fail. Money shrinks, and researchers try new roads.", "humanExplainDisplay": "AI Winter is the day after\na ==school talent show flops==.\nThe ==fog machine== gets packed away,\nand the pizza money disappears.\n\nIt explains AI’s cold spells\nafter big promises fail.\nMoney shrinks,\nand researchers try new roads.", "relationsNarrative": "Dartmouth Workshop\nThe Dartmouth Workshop sparked early AI hope, and AI Winter came when that hope fell short.\n\nExpert System\nThe Expert System boom broke and helped trigger the second AI Winter.\n\nAI Bubble\nAn AI Bubble is the hot part. AI Winter is the quiet part after it pops.\n\nDeep Learning\nDeep Learning used better results and more computer power to pull AI back into the spotlight.", "relations": { "dartmouth-workshop": { "label": "follows the hopes from …", "note": "Early AI optimism made the later letdown feel huge." }, "expert-system": { "label": "followed the bust of …", "note": "The Expert System boom broke and helped start the second AI Winter." }, "ai-bubble": { "label": "comes after … bursts", "note": "When an AI Bubble pops, an AI Winter often follows." }, "deep-learning": { "label": "warmed up by …", "note": "Deep Learning helped AI leave the cold spell." } } }, "zh": { "fullName": "AI 寒冬", "factExplain": "AI 研究因预期落空而遭遇资金与关注骤减的时期。", "humanExplain": "AI 寒冬是网红奶茶退潮：吹成必喝，卖不动后投资人撤灯牌。\n\n用来理解泡沫退潮后，资金收缩与研究路线转向。", "humanExplainDisplay": "AI 寒冬是\n==网红奶茶退潮==：\n吹成必喝，\n卖不动后投资人撤灯牌。\n\n用来理解泡沫退潮后，\n资金收缩，\n与研究路线转向。", "relationsNarrative": "Dartmouth Workshop\n它点燃早期 AI 乐观，寒冬是预期落空后的反噬。\n\nExpert System\n专家系统热潮破裂，是第二次 AI 寒冬的重要导火索。\n\nAI Bubble\n泡沫讲的是过热，寒冬讲的是过热退潮后的冷清。\n\nDeep Learning\n深度学习用效果和算力，把 AI 从低谷拉回聚光灯。", "relations": { "dartmouth-workshop": { "label": "承接…的早期热望", "note": "早期乐观预期埋下巨大落差。" }, "expert-system": { "label": "见证…退潮", "note": "专家系统泡沫破裂触发第二次寒冬。" }, "ai-bubble": { "label": "对应…破裂", "note": "泡沫退潮后，寒冬常接着到来。" }, "deep-learning": { "label": "被…重新升温", "note": "深度学习让 AI 再次走出低谷。" } } } } }, { "id": "ai-workforce-transition", "name": "AI Workforce Transition", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "automation-job" }, { "to": "ai-career-moat" }, { "to": "ai-literacy-ai" }, { "to": "ai-native-organization" } ], "track": "today", "i18n": { "en": { "fullName": "AI Workforce Transition", "factExplain": "The shift in jobs and work roles caused by AI.", "humanExplain": "AI workforce transition is like the school cafeteria hiding the pizza line. The lunchroom is the same, but the old route gets you peas.\n\nIt changes hiring, training, and job design. Companies need a new plan, and workers do too.", "humanExplainDisplay": "AI workforce transition is like\nthe school cafeteria ==hiding the pizza line==.\nThe lunchroom is the same,\nbut the ==old route gets you peas==.\n\nIt changes hiring, training, and job design.\nCompanies need a new plan,\nand workers do too.", "relationsNarrative": "Automation-job\nAI workforce transition often starts when specific tasks get automated.\n\nAI career moat\nIt pushes people to build skills AI cannot easily replace.\n\nAI Literacy\nAI literacy helps workers keep up as roles change.\n\nAI-native organization\nAI-native organizations redesign work around humans and AI.", "relations": { "automation-job": { "label": "brings changes in …", "note": "Automation replaces or reshapes some tasks." }, "ai-career-moat": { "label": "pushes people to build …", "note": "A moat comes from skills AI cannot easily replace." }, "ai-literacy-ai": { "label": "requires stronger …", "note": "AI literacy helps people move into new roles." }, "ai-native-organization": { "label": "leads to …", "note": "Work gets redesigned around humans and AI working together." } } }, "zh": { "fullName": "AI 劳动力转型", "factExplain": "AI 推动岗位需求和工作分工重组的过程。", "humanExplain": "AI 劳动力转型像地铁改线：站还在，换乘规则变了，不看新图的人先绕晕。\n\n影响招聘、培训和岗位设计，企业和个人都要重规划。", "humanExplainDisplay": "AI 劳动力转型像==地铁改线==：\n站还在，\n==换乘规则==变了，\n不看新图的人先绕晕。\n\n影响招聘、培训和岗位设计，\n企业和个人都要重规划。", "relationsNarrative": "Automation-job\nAI 转型常从自动化具体任务开始。\n\nAI Career Moat\n它迫使个人寻找更难被替代的能力。\n\nAI Literacy\nAI 素养决定员工能否跟上岗位变化。\n\nAI-native Organization\n组织形态变化会重新设计人机分工。", "relations": { "automation-job": { "label": "带来…变化", "note": "自动化让部分任务被替换或重组。" }, "ai-career-moat": { "label": "倒逼建立…", "note": "护城河来自难被自动化的能力。" }, "ai-literacy-ai": { "label": "要求提升…", "note": "AI 素养决定谁能顺利转岗。" }, "ai-native-organization": { "label": "催生…", "note": "组织会围绕人机协作重新分工。" } } } } }, { "id": "ai-worm", "name": "AI Worm", "layer": "L6", "era": "2024", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "prompt-injection" }, { "to": "agent-security" }, { "to": "data-exfiltration" }, { "to": "agent-internet" } ], "track": "today", "i18n": { "en": { "fullName": "AI Worm", "factExplain": "A harmful program that uses AI tools to copy and spread itself.", "humanExplain": "An AI Worm is like a scam text in a group chat. It tricks one AI helper, then makes it DM the next one.\n\nIt can hit online agents, email, and plugins. It may steal data or send bad commands.", "humanExplainDisplay": "An AI Worm is like a ==scam text==\nin a group chat.\nIt tricks one AI helper,\nthen makes it ==DM the next one==.\n\nIt can hit online agents,\nemail,\nand plugins.\nIt may steal data\nor send bad commands.", "relationsNarrative": "Prompt injection\nAn AI Worm can use a bad prompt to make a model spread it.\n\nAgent Security\nAn AI Worm shows why online agents need strong safety checks.\n\nExfiltration\nAn AI Worm may steal private data while it spreads.\n\nAgent Internet\nAn AI Worm spreads more easily when many agents connect to each other.", "relations": { "prompt-injection": { "label": "gives orders through …", "note": "A bad prompt can trick a model into spreading the worm." }, "agent-security": { "label": "tests …", "note": "AI Worms are a clear danger for agents that go online." }, "data-exfiltration": { "label": "often comes with …", "note": "The worm may steal private data as it spreads." }, "agent-internet": { "label": "spreads through …", "note": "More connected agents give the worm more paths to travel." } } }, "zh": { "fullName": "AI 蠕虫", "factExplain": "能借 AI 工具链自我传播的恶意程序。", "humanExplain": "AI 蠕虫像家族群诈骗链接：骗到一个助手，就让它转发坑下家。\n\n它威胁联网代理、邮箱和插件，可偷数据、乱发指令。", "humanExplainDisplay": "AI 蠕虫像家族群诈骗链接：\n==骗到一个助手==，\n就让它转发，\n==坑下家==。\n\n它威胁联网代理、邮箱和插件，\n可偷数据，\n乱发指令。", "relationsNarrative": "Prompt injection\nAI 蠕虫常借恶意提示，让模型替它转发。\n\nAgent Security\n它暴露代理能上网、能行动后的安全短板。\n\nExfiltration\n传播过程中，它可能顺手偷走敏感数据。\n\nAgent Internet\n代理彼此连接越多，蠕虫越容易扩散。", "relations": { "prompt-injection": { "label": "借…下指令", "note": "恶意提示可诱导模型转发蠕虫。" }, "agent-security": { "label": "考验…", "note": "它是联网代理的典型安全威胁。" }, "data-exfiltration": { "label": "常伴随…", "note": "传播时可能顺手带走敏感数据。" }, "agent-internet": { "label": "在…中扩散", "note": "代理互联让传播链条更长。" } } } } }, { "id": "ai-youth-safety", "name": "AI Youth Safety", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "ai-companion-risk" }, { "to": "ai-tutor" }, { "to": "ai-literacy-ai" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "AI Youth Safety", "factExplain": "Rules and design choices that help protect kids from AI harm.", "humanExplain": "AI Youth Safety is a bike helmet for AI. Kids can still ride, but not off the garage roof.\n\nIt sets age limits, blocks risky replies, and gives clear warnings. You meet it in AI tutors, AI friend apps, and teen chat apps.", "humanExplainDisplay": "AI Youth Safety is a ==bike helmet== for AI.\nKids can still ride,\nbut not ==off the garage roof==.\n\nIt sets age limits,\nblocks risky replies,\nand gives clear warnings.\nYou meet it in AI tutors,\nAI friend apps,\nand teen chat apps.", "relationsNarrative": "Companion-risk\nAI Youth Safety reduces the risk of kids depending on AI companions too much.\n\nAI Tutor\nAI tutors need a balance between good learning and strong protection.\n\nAI Literacy\nAI Literacy helps teens spot pressure, bias, and false advice.\n\nAI-regulation\nAI Youth Safety often uses age ratings and platform duties from regulation.", "relations": { "ai-companion-risk": { "label": "reduces …", "note": "Kids can depend too much on human-like AI friends." }, "ai-tutor": { "label": "sets guardrails for …", "note": "AI tutors must teach well and keep kids safe." }, "ai-literacy-ai": { "label": "builds self-protection with …", "note": "AI Literacy helps teens spot pressure, bias, and bad advice." }, "ai-regulation": { "label": "relies on …", "note": "Rules make age ratings and platform duties real." } } }, "zh": { "fullName": "青少年 AI 安全", "factExplain": "保护未成年人使用 AI 时免受伤害的治理实践。", "humanExplain": "青少年 AI 安全像游乐园手环：不是不让玩，是挡住不该坐的过山车。\n\n用于教育、陪伴和社交产品，管分级、过滤和告知。", "humanExplainDisplay": "青少年 AI 安全像游乐园手环：\n不是不让玩，\n是挡住==不该坐的过山车==。\n\n用于教育、陪伴和社交产品，\n管分级、过滤和告知。", "relationsNarrative": "Companion-risk\n它重点降低未成年人对 AI 陪伴的依赖风险。\n\nAI Tutor\n家教类 AI 需要在学习效果和保护之间找平衡。\n\nAI Literacy\nAI 素养让青少年更会识别诱导、偏见和误导。\n\nAI-regulation\n青少年保护常靠年龄分级和平台责任落地。", "relations": { "ai-companion-risk": { "label": "降低…", "note": "未成年人更易依赖拟人化陪伴。" }, "ai-tutor": { "label": "约束…使用", "note": "教育场景要兼顾效果与保护。" }, "ai-literacy-ai": { "label": "用…提升自护", "note": "会用 AI，也要会识别诱导。" }, "ai-regulation": { "label": "借…落地", "note": "年龄分级和平台责任靠规则。" } } } } }, { "id": "ai", "name": "AI Companion", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "chatbot" }, { "to": "personal-ai-apps" }, { "to": "affective-computing" }, { "to": "ai-companion-risk" } ], "track": "today", "seo": { "zh": { "title": "AI 搭子 / AI 陪伴 是什么?深夜便利店搭子,一文看懂 — AI Rookies", "description": "面向陪伴与协作的个人化 AI 应用。人话版一分钟讲明白。" }, "en": { "title": "What Is AI Companion? The 2 AM Chat Buddy", "description": "A personal AI made to keep you company and help you work. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "AI Companion", "factExplain": "A personal AI made to keep you company and help you work.", "humanExplain": "An AI Companion is like a group chat buddy who always texts back. Helpful at 2 a.m., but not your real best friend.\n\nYou meet it in chat apps and personal AI tools. It can support your mood or nudge study, but it is not a real person.", "humanExplainDisplay": "An AI Companion is like a ==group chat buddy==\nwho ==always texts back==.\nHelpful at 2 a.m.,\nbut not your real best friend.\n\nYou meet it in chat apps\nand personal AI tools.\nIt can support your mood\nor nudge study,\nbut it is not a real person.", "relationsNarrative": "Chatbot\nAn AI Companion often appears as a chatbot.\n\nPersonal AI apps\nIt is the companionship branch of personal AI apps.\n\nAffective Computing\nAffective Computing helps it read tone and mood.\n\nCompanion-risk\nLong-term reliance can blur the line between AI and people.", "relations": { "chatbot": { "label": "turns … into company", "note": "It moves chatbots from quick answers to long-term company." }, "personal-ai-apps": { "label": "is a type of …", "note": "It is the emotional side of personal AI apps." }, "affective-computing": { "label": "uses … to read mood", "note": "Affective Computing helps it notice tone and feelings." }, "ai-companion-risk": { "label": "may cause …", "note": "Too much reliance can blur the line between AI and people." } } }, "zh": { "fullName": "AI 搭子 / AI 陪伴", "factExplain": "面向陪伴与协作的个人化 AI 应用。", "humanExplain": "AI 搭子像深夜便利店店员：你随时进来唠两句，它回应，但不是家人。\n\n用于聊天陪伴、学习督促和情绪支持，但要守住人机边界。", "humanExplainDisplay": "AI 搭子像\n==深夜便利店店员==：\n你随时进来唠两句，\n它回应，但不是家人。\n\n用于聊天陪伴、学习督促\n和情绪支持，\n但要守住人机边界。", "relationsNarrative": "Chatbot\nAI 搭子常以聊天机器人形态出现。\n\nPersonal AI Apps\n它是个人 AI 应用里的陪伴型分支。\n\nAffective Computing\n情感计算帮助它识别语气和情绪。\n\nCompanion-risk\n长期依赖可能模糊人机关系边界。", "relations": { "chatbot": { "label": "升级…为陪伴", "note": "从问答聊天走向长期陪伴。" }, "personal-ai-apps": { "label": "属于…的一类", "note": "它是个人 AI 的情感型入口。" }, "affective-computing": { "label": "借…识别情绪", "note": "情绪识别让回应更像懂你。" }, "ai-companion-risk": { "label": "可能带来…", "note": "过度依赖会模糊关系边界。" } } } } }, { "id": "alexnet", "name": "AlexNet", "layer": "L3", "era": "2012", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "deep-learning" }, { "to": "resnet" }, { "to": "gpu" } ], "track": "history", "i18n": { "en": { "fullName": "AlexNet Convolutional Neural Network", "factExplain": "A deep CNN that crushed image recognition scores in 2012.", "humanExplain": "AlexNet was the kid with a jetpack at a bike race. The old image tools were still pedaling.\n\nIt showed deep networks could understand pictures far better. It made computer vision hot, then pushed deep learning into real products.", "humanExplainDisplay": "AlexNet was the ==kid with a jetpack==\nat a bike race.\nThe old image tools were ==still pedaling==.\n\nIt showed deep networks could understand pictures far better.\nIt made computer vision hot,\nthen pushed deep learning into real products.", "relationsNarrative": "CNN\nAlexNet is a classic early CNN for image recognition.\n\nDeep Learning\nAlexNet helped turn deep learning from a niche idea into a mainstream tool.\n\nResNet\nResNet built on the deeper vision path that AlexNet made famous.\n\nGPU\nAlexNet relied on GPUs for training and helped raise demand for more compute.", "relations": { "cnn": { "label": "is a landmark …", "note": "AlexNet is an early landmark CNN for image recognition." }, "deep-learning": { "label": "helped bring … mainstream", "note": "AlexNet helped move deep learning from labs to real products." }, "resnet": { "label": "paved the way for …", "note": "ResNet kept pushing the deeper vision path AlexNet made famous." }, "gpu": { "label": "trained on …", "note": "AlexNet needed GPUs to train well and fast." } } }, "zh": { "fullName": "AlexNet 卷积神经网络", "factExplain": "2012 年大幅刷新图像识别成绩的深度卷积网络。", "humanExplain": "AlexNet 像高考考场里突然杀出的黑马，一下把旧题路数全打懵：原来网络堆深了，真能把看图看明白。\n\n它带火了计算机视觉，也让深度学习从学术圈快速走向工业落地。", "humanExplainDisplay": "AlexNet 像高考考场里\n突然杀出的==黑马==，\n一下把旧题路数全打懵：\n原来网络堆深了，\n真能把==看图看明白==。\n\n它带火了计算机视觉，\n也让深度学习从学术圈\n快速走向工业落地。", "relationsNarrative": "Cnn\n它是卷积神经网络的经典代表作之一。\n\nDeep-learning\n它的成功让深度学习从冷门方向变成主流。\n\nResNet\nResNet 延续并升级了它开创的深层视觉路线。\n\nGPU\n它依赖 GPU 训练，也带动了算力需求上升。", "relations": { "cnn": { "label": "属于…代表作", "note": "它是卷积网络早期里程碑。" }, "deep-learning": { "label": "带火…落地", "note": "它让深度学习开始大规模出圈。" }, "resnet": { "label": "为…铺路", "note": "后来的更深视觉模型接着往前走。" }, "gpu": { "label": "依赖…训练", "note": "没有 GPU，很难把它高效训出来。" } } } } }, { "id": "alignment", "name": "Alignment", "layer": "L6", "era": "2016", "publishedAt": "2026-05-23T11:40:00Z", "relations": [ { "to": "rlhf" }, { "to": "agi" }, { "to": "ai-bias" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "Alignment", "factExplain": "Making AI goals and actions fit human intent and values.", "humanExplain": "Alignment is teaching a super-fast robot chef the house rules. Dinner is great, but a flaming kitchen is not extra credit.\n\nIt checks if AI understands the goal and respects limits. It matters more as AI gets stronger.", "humanExplainDisplay": "Alignment is teaching a ==super-fast robot chef==\nthe house rules.\nDinner is great,\nbut a ==flaming kitchen== is not extra credit.\n\nIt checks if AI understands the goal\nand respects limits.\nIt matters more as AI gets stronger.", "relationsNarrative": "RLHF\nRLHF uses human feedback to move models toward Alignment.\n\nAGI\nAGI makes Alignment harder and more important.\n\nAI-bias\nAI-bias can be a real sign of weak Alignment.\n\nAI-regulation\nAI-regulation adds outside rules to support Alignment.", "relations": { "rlhf": { "label": "learns from …", "note": "RLHF uses human feedback to steer models toward Alignment." }, "agi": { "label": "matters more with …", "note": "AGI makes Alignment harder and more important." }, "ai-bias": { "label": "can show as …", "note": "AI-bias can show up when Alignment is weak." }, "ai-regulation": { "label": "is backed by …", "note": "AI-regulation adds rules and pressure around Alignment." } } }, "zh": { "fullName": "对齐", "factExplain": "让 AI 的目标和行为符合人类意图与价值的研究方向。", "humanExplain": "对齐像给代驾讲规矩：车技再好，也得把你送回家，不能顺路去撸串。\n\n它用于训练、评测和治理模型，降低失控、作恶或误解指令的风险。", "humanExplainDisplay": "对齐像==给代驾讲规矩==：\n车技再好，\n也得把你送回家，\n不能==顺路去撸串==。\n\n它用于训练、评测和治理模型，\n降低失控、作恶或误解指令的风险。", "relationsNarrative": "RLHF\nRLHF 通过人类反馈让模型更接近 Alignment 目标。\n\nAGI\nAGI 提升了 Alignment 问题的复杂度和重要性。\n\nAI-bias\nAI-bias 是 Alignment 不充分时的现实表现之一。\n\nAI-regulation\nAI-regulation 从制度层面对 Alignment 形成补充。", "relations": { "rlhf": { "label": "靠…实现" }, "agi": { "label": "因…更重要" }, "ai-regulation": { "label": "被…强调" } } } } }, { "id": "alpha-beta-pruning", "name": "Α-β Pruning", "layer": "L2", "era": "1958", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "monte-carlo-tree-search" }, { "to": "heuristic-search" }, { "to": "deep-q-network" }, { "to": "automated-planning" } ], "track": "history", "i18n": { "en": { "fullName": "Alpha-Beta Pruning", "factExplain": "A shortcut that skips losing branches during minimax game search.", "humanExplain": "Alpha-Beta Pruning is like playing chess with a strict grandma. Once your move is already losing, she stops watching your “genius” plan.\n\nGame AIs use it to search moves faster. It skips branches but keeps the same best move.", "humanExplainDisplay": "Alpha-Beta Pruning is like playing chess\nwith a ==strict grandma==.\nOnce your move is already losing,\nshe stops watching your ==“genius” plan==.\n\nGame AIs use it to search moves faster.\nIt skips branches\nbut keeps the same best move.", "relationsNarrative": "MCTS\nBoth search decision trees, but Alpha-Beta cuts branches while MCTS samples plays.\n\nHeuristic Search\nGood move order helps Alpha-Beta cut branches earlier.\n\nDQN\nAlpha-Beta uses clear tree search. DQN uses learned value guesses.\n\nPlanning\nAlpha-Beta is a classic search trick in planning and game solving.", "relations": { "monte-carlo-tree-search": { "label": "contrasts with … search", "note": "Alpha-Beta prunes branches. MCTS samples trial games." }, "heuristic-search": { "label": "speeds up with …", "note": "Good move order helps it cut more branches." }, "deep-q-network": { "label": "differs from … decisions", "note": "It searches a tree. DQN uses learned values." }, "automated-planning": { "label": "belongs to classic …", "note": "It is a classic search trick in planning and games." } } }, "zh": { "fullName": "Alpha-Beta 剪枝", "factExplain": "一种在极小化极大搜索中提前跳过无用分支的方法。", "humanExplain": "像武侠过招先识破败势：这一路注定要输，后面再花哨的招，也没必要陪它演完。\n\n常用于博弈程序搜索走法，在不改答案的前提下大幅少算分支。", "humanExplainDisplay": "像武侠过招==先识破败势==：\n这一路注定要输，\n后面再花哨的招，\n也==没必要陪它演完==。\n\n常用于博弈程序搜索走法，\n在不改答案的前提下，\n大幅少算分支。", "relationsNarrative": "Monte-Carlo-Tree-Search\n两者都搜决策树，但一个靠剪枝，一个靠采样。\n\nHeuristic-Search\n好的启发式排序能让它更早剪枝、更省计算。\n\nDQN\n它是显式树搜索，DQN 则用学习到的价值决策。\n\nAutomated-Planning\n它属于经典搜索派，常见于规划和博弈求解。", "relations": { "monte-carlo-tree-search": { "label": "对比…搜索", "note": "一个靠剪枝推演，一个靠采样试探。" }, "heuristic-search": { "label": "常配合…提速", "note": "好启发顺序能让它剪掉更多分支。" }, "deep-q-network": { "label": "区别于…决策", "note": "它靠显式搜索，不靠神经网络估值。" }, "automated-planning": { "label": "属于…经典方法", "note": "是早期规划与博弈里的代表技巧。" } } } } }, { "id": "alphafold-2", "name": "AlphaFold 2", "layer": "L4", "era": "2020", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "ai-for-science" }, { "to": "ai-drug-discovery" }, { "to": "attention" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "AlphaFold 2 Protein Structure Prediction Model", "factExplain": "A deep learning model for predicting a protein’s 3D shape.", "humanExplain": "AlphaFold 2 is like a genius kid with a bead bracelet. Give it the bead order. It pictures the twisty 3D shape.\n\nScientists use it in biology and drug research. It cuts down some lab trial and error.", "humanExplainDisplay": "AlphaFold 2 is like a ==genius kid==\nwith a bead bracelet.\nGive it the ==bead order==.\nIt pictures the twisty 3D shape.\n\nScientists use it in biology and drug research.\nIt cuts down some lab trial and error.", "relationsNarrative": "AI for Science\nAlphaFold 2 is a landmark result in AI for Science.\n\nAI Drug Discovery\nProtein shape prediction helps drug teams find targets faster.\n\nAttention\nAlphaFold 2 uses Attention to model links between amino acids.\n\nDeep Learning\nAlphaFold 2 uses Deep Learning to learn from many known structures.", "relations": { "ai-for-science": { "label": "marks a breakthrough in …", "note": "It put AI on the main stage of research." }, "ai-drug-discovery": { "label": "speeds up …", "note": "Protein shapes can help teams study drug targets sooner." }, "attention": { "label": "models links with …", "note": "Attention helps it track how amino acids affect each other." }, "deep-learning": { "label": "is built on …", "note": "Deep Learning lets it learn patterns from many known structures." } } }, "zh": { "fullName": "蛋白质结构预测模型 AlphaFold 2", "factExplain": "预测蛋白质三维结构的深度学习模型。", "humanExplain": "AlphaFold 2 像折纸高手看纸痕：只凭氨基酸清单，就折出蛋白质立体形。\n\n用于生物研究和药物发现，减少湿实验试错。", "humanExplainDisplay": "AlphaFold 2 像\n==折纸高手==看纸痕：\n只凭氨基酸清单，\n就折出蛋白质立体形。\n\n用于生物研究和药物发现，\n减少湿实验试错。", "relationsNarrative": "AI For Science\nAlphaFold 2 是 AI for Science 的标志性成果。\n\nAI Drug Discovery\n结构预测能帮助药物发现更快锁定靶点。\n\nAttention\n它用注意力机制建模氨基酸之间的关系。\n\nDeep Learning\n它依靠深度学习从海量结构数据中学习规律。", "relations": { "ai-for-science": { "label": "代表…突破", "note": "它把 AI 推上科研主舞台。" }, "ai-drug-discovery": { "label": "加速…", "note": "结构预测能缩短靶点研究前期。" }, "attention": { "label": "借助…建模关系", "note": "它用注意力理解氨基酸互作。" }, "deep-learning": { "label": "基于…", "note": "端到端学习替代大量手工规则。" } } } } }, { "id": "alphago", "name": "AlphaGo", "layer": "L4", "era": "2016", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "monte-carlo-tree-search" }, { "to": "deep-reinforcement-learning" }, { "to": "alphazero" } ], "track": "history", "i18n": { "en": { "fullName": "AlphaGo", "factExplain": "A Go-playing AI built with deep learning and reinforcement learning.", "humanExplain": "AlphaGo turns game night into boot camp. It plays itself all night, then makes Go champions sweat.\n\nIt used Deep RL to master Go. It made RL famous outside labs. It showed AI could plan tough moves.", "humanExplainDisplay": "AlphaGo turns ==game night== into ==boot camp==.\nIt plays itself all night,\nthen makes Go champions sweat.\n\nIt used Deep RL to master Go.\nIt made RL famous outside labs.\nIt showed AI could plan tough moves.", "relationsNarrative": "RL\nAlphaGo used repeated games and feedback to improve its strategy.\n\nMCTS\nAlphaGo used MCTS to judge positions and choose stronger moves.\n\nDeep RL\nAlphaGo became the breakout example of Deep RL.\n\nAlphaZero\nAlphaZero built on AlphaGo and spread the idea to more board games.", "relations": { "reinforcement-learning": { "label": "gets stronger with …", "note": "AlphaGo learned better moves by playing many games." }, "monte-carlo-tree-search": { "label": "searches moves with …", "note": "MCTS helped AlphaGo judge the next move." }, "deep-reinforcement-learning": { "label": "is a milestone in …", "note": "AlphaGo made Deep RL famous outside research labs." }, "alphazero": { "label": "was generalized by …", "note": "AlphaZero took the idea to more board games." } } }, "zh": { "fullName": "阿尔法围棋", "factExplain": "一个用深度学习下围棋的强化学习模型。", "humanExplain": "它像刷题刷到封神的竞赛生：先把历年真题啃透，再自己疯狂模考，最后连老师都得倒吸一口气。\n\n它让强化学习真正出圈，也带火了 AI 在博弈、规划和决策里的影响力。", "humanExplainDisplay": "它像刷题刷到\n==封神==的竞赛生：\n先把历年真题啃透，\n再自己疯狂==模考==，\n最后连老师都得\n倒吸一口气。\n\n它让强化学习真正出圈，\n也带火了 AI 在博弈、\n规划和决策里的\n影响力。", "relationsNarrative": "Reinforcement-learning\nAlphaGo 通过反复对弈和反馈更新策略。\n\nMonte-carlo-tree-search\n它用树搜索评估局面，并选择更优落子。\n\nDeep-reinforcement-learning\nAlphaGo 是深度强化学习出圈的标志性案例。\n\nAlphazero\nAlphaZero 延续其思路，并推广到更多棋类。", "relations": { "reinforcement-learning": { "label": "靠…持续变强", "note": "它靠反复对弈学习更优策略。" }, "monte-carlo-tree-search": { "label": "用…搜索落子", "note": "它靠树搜索评估下一手。" }, "deep-reinforcement-learning": { "label": "属于…里程碑", "note": "它是深度强化学习代表作。" }, "alphazero": { "label": "被…进一步泛化", "note": "后者把方法扩展到更多棋类。" } } } } }, { "id": "alphazero", "name": "AlphaZero", "layer": "L4", "era": "2017", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "alphago" }, { "to": "reinforcement-learning" }, { "to": "monte-carlo-tree-search" }, { "to": "deep-reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "AlphaZero", "factExplain": "A reinforcement learning model that learns board-game strategy by playing itself.", "humanExplain": "AlphaZero is the kid who plays both sides of the chessboard at lunch. It beats itself, gloats a little, then comes back stronger.\n\nIt starts with the rules, then learns by self-play. It is famous for Go, chess, and shogi.", "humanExplainDisplay": "AlphaZero is the kid who ==plays both sides==\nof the chessboard at lunch.\nIt ==beats itself==, gloats a little,\nthen comes back stronger.\n\nIt starts with the rules,\nthen learns by self-play.\nIt is famous for Go, chess, and shogi.", "relationsNarrative": "AlphaGo\nAlphaZero is a more general AlphaGo, built for more than Go.\n\nRL\nAlphaZero improves by playing itself and learning from wins and losses.\n\nMCTS\nAlphaZero uses MCTS to search ahead and pick stronger moves.\n\nDeep RL\nAlphaZero is one famous example of Deep RL.", "relations": { "alphago": { "label": "upgrades …", "note": "It moves from Go specialist to many-game player." }, "reinforcement-learning": { "label": "learns through …", "note": "RL improves it with rewards from wins and losses." }, "monte-carlo-tree-search": { "label": "chooses moves with …", "note": "MCTS helps it search ahead before each move." }, "deep-reinforcement-learning": { "label": "shows the power of …", "note": "It is a famous win for Deep RL." } } }, "zh": { "fullName": "AlphaZero", "factExplain": "一种靠自我对弈学习棋类策略的强化学习模型。", "humanExplain": "像健身房里对着镜子加练的人，没人手把手带，它自己跟自己较劲，最后把动作卷成标准答案。\n\n主要用于围棋、国际象棋和将棋，证明 AI 靠自我对弈也能学出顶级策略。", "humanExplainDisplay": "像健身房里\n对着镜子加练的人，\n没人手把手带，\n它自己跟自己==较劲==，\n最后把动作卷成\n==标准答案==。\n\n主要用于围棋、\n国际象棋和将棋，\n证明 AI 靠自我对弈\n也能学出顶级策略。", "relationsNarrative": "Alphago\n它可看作 AlphaGo 的通用化升级，不再只为围棋设计。\n\nReinforcement-learning\n它靠自我对弈和胜负反馈，持续改进下棋策略。\n\nMonte-carlo-tree-search\n它在对局时配合 MCTS 搜索，更高效挑选落子。\n\nDeep-reinforcement-learning\n它是深度强化学习的代表成果之一。", "relations": { "alphago": { "label": "是…的升级版", "note": "从围棋专才走向多棋类通才。" }, "reinforcement-learning": { "label": "属于…范式", "note": "核心是靠奖励信号自我提升。" }, "monte-carlo-tree-search": { "label": "结合…找好棋", "note": "搜索帮助它在对局中挑行动。" }, "deep-reinforcement-learning": { "label": "是…代表作", "note": "把深度网络与强化学习结合。" } } } } }, { "id": "answer-engine-optimization", "name": "AEO", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "agentic-search" }, { "to": "rag" }, { "to": "information-retrieval" }, { "to": "content-provenance" } ], "track": "today", "i18n": { "en": { "fullName": "Answer Engine Optimization", "factExplain": "A method for making content more likely to be cited by AI answers.", "humanExplain": "AEO makes your page the kid with color-coded notes. AI needs an answer, so it borrows from the neat binder.\n\nYou use it on websites and help centers. News sites use it too, to get cited by AI answers.", "humanExplainDisplay": "AEO makes your page the kid with ==color-coded notes==.\nAI needs an answer,\nso it borrows from the ==neat binder==.\n\nYou use it on websites and help centers.\nNews sites use it too,\nto get cited by AI answers.", "relationsNarrative": "Agentic search\nAEO tries to get your content mentioned in AI search answers.\n\nRAG\nClear content is easier for RAG to find and feed to the model.\n\nIR\nIR still decides if the system finds your content first.\n\nContent provenance\nClear sources make answer engines more willing to cite you.", "relations": { "agentic-search": { "label": "seeks exposure in …", "note": "AEO aims to appear inside direct AI search answers." }, "rag": { "label": "fits … source needs", "note": "Clear, retrievable content is easier for RAG to use." }, "information-retrieval": { "label": "gets found through …", "note": "IR quality decides if the content is seen first." }, "content-provenance": { "label": "builds trust with …", "note": "Clear sources make answer engines safer to cite." } } }, "zh": { "fullName": "答案引擎优化", "factExplain": "优化内容以被 AI 答案引用的方法。", "humanExplain": "AEO像替内容上相亲节目：不只露脸，还得让AI红娘当场夸你靠谱。\n\n用于官网、知识库和媒体，提高被答案引用概率。", "humanExplainDisplay": "AEO像替内容上\n==相亲节目==：\n不只露脸，\n还得让==AI红娘==当场夸你靠谱。\n\n用于官网、知识库和媒体，\n提高被答案引用概率。", "relationsNarrative": "Agentic Search\n答案优化的目标，是在 AI 搜索回答里被提到。\n\nRAG\n内容结构清楚，才更容易被检索后喂给模型。\n\nInformation Retrieval\n传统检索仍决定内容能否先被系统找到。\n\nContent Provenance\n来源越清楚，答案引擎越敢引用。", "relations": { "agentic-search": { "label": "面向…争曝光", "note": "答案优化瞄准 AI 搜索的直接回答。" }, "rag": { "label": "适配…资料入口", "note": "可检索的内容更容易进入生成答案。" }, "information-retrieval": { "label": "借…被找到", "note": "检索质量决定内容能否先被看见。" }, "content-provenance": { "label": "强化…可信度", "note": "来源清楚，答案引擎才更敢引用。" } } } } }, { "id": "anti-distillation", "name": "Anti-distillation", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "distillation" }, { "to": "open-weights" }, { "to": "api" }, { "to": "copyright" } ], "track": "today", "i18n": { "en": { "fullName": "Anti-distillation", "factExplain": "Methods used to stop others from copying a model’s skills.", "humanExplain": "Anti-distillation is the secret-sauce move. You can taste the burger, but the kitchen keeps the recipe.\n\nIt protects closed models and paid APIs. It makes mass copying harder.", "humanExplainDisplay": "Anti-distillation is the ==secret-sauce== move.\nYou can taste the burger,\nbut the kitchen keeps the ==recipe==.\n\nIt protects closed models and paid APIs.\nIt makes mass copying harder.", "relationsNarrative": "Distillation\nAnti-distillation tries to stop distillation from copying a model.\n\nOpen weights\nOpen weights make anti-distillation much harder.\n\nAPI\nMany anti-distillation steps use API limits and output rules.\n\nCopyright\nAnti-distillation blocks copying, and copyright rules handle misuse.", "relations": { "distillation": { "label": "blocks … copying", "note": "It is the defense against model distillation." }, "open-weights": { "label": "is weaker with …", "note": "Open weights make copying much harder to stop." }, "api": { "label": "often runs through …", "note": "Many defenses live at the API layer." }, "copyright": { "label": "supports …", "note": "Technical blocks and copyright rules often work together." } } }, "zh": { "fullName": "反蒸馏", "factExplain": "防止模型被模仿或提取能力的技术与策略。", "humanExplain": "像武侠门派防人偷学心法：你能在擂台上看几招，但内功怎么运、杀招怎么接，不会让你全摸走。\n\n常用于保护闭源模型和商业 API，减少被批量模仿、复制能力的风险。", "humanExplainDisplay": "像武侠门派防人偷学==心法==：\n你能在擂台上看几招，\n但内功怎么运、杀招怎么接，\n不会让你全==摸走==。\n\n常用于保护闭源模型和商业 API，\n减少被批量模仿、\n复制能力的风险。", "relationsNarrative": "Distillation\n它的目标，就是降低模型被蒸馏复制的可能。\n\nOpen-weights\n模型权重一旦开放，反蒸馏就更难发挥作用。\n\nAPI\n很多反蒸馏措施通过接口限流和输出设计实现。\n\nCopyright\n技术防护挡复制，版权规则管滥用，两者常配合。", "relations": { "distillation": { "label": "对抗…复制", "note": "它就是针对模型蒸馏的防护思路。" }, "open-weights": { "label": "与…相对", "note": "权重越开放，越难靠它阻止模仿。" }, "api": { "label": "常经由…实施", "note": "很多防护做在接口层而非模型层。" }, "copyright": { "label": "补充…保护", "note": "技术防护常与法律约束一起用。" } } } } }, { "id": "apertus", "name": "Apertus", "layer": "L3", "era": "2025", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "open-source-model" }, { "to": "open-weights" }, { "to": "multilingual-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Swiss Open Large Language Model", "factExplain": "An open-weight LLM released by Swiss institutions.", "humanExplain": "Apertus is like a Lego robot kit with the instructions still in the box. You can turn it on, then rebuild it your way.\n\nIt is useful for research, private setups, and local-language apps. The key point is simple: people can check it and repeat the work.", "humanExplainDisplay": "Apertus is like a ==Lego robot kit==\nwith the instructions still in the box.\nYou can turn it on,\nthen ==rebuild it your way==.\n\nIt is useful for research,\nprivate setups,\nand local-language apps.\nThe key point is simple:\npeople can check it\nand repeat the work.", "relationsNarrative": "LLM\nApertus is an LLM for understanding and writing text.\n\nOpen-source-model\nApertus follows the open model path for study and change.\n\nOpen weights\nOpen weights let users download and run Apertus themselves.\n\nMultilingual AI\nApertus focuses on many languages, so more people can use it.", "relations": { "llm": { "label": "is a kind of …", "note": "Apertus is a large text model that reads and writes text." }, "open-source-model": { "label": "follows the … path", "note": "Its open release helps people study and change it." }, "open-weights": { "label": "releases …", "note": "Open weights make private deployment possible." }, "multilingual-ai": { "label": "targets …", "note": "Apertus is built for use across many languages." } } }, "zh": { "fullName": "瑞士开放大语言模型", "factExplain": "瑞士机构发布的开放权重大语言模型。", "humanExplain": "Apertus 像把模型样板房连图纸公开：不只让你住，还能照着改造。\n\n适合研究、私有部署、本地化应用，重点是可审计、可复现。", "humanExplainDisplay": "Apertus 像把==模型样板房==\n连图纸公开：\n不只让你住，\n还能照着改造。\n\n适合研究、私有部署、\n本地化应用，\n重点是可审计、可复现。", "relationsNarrative": "LLM\n它本质上是能理解并生成文本的大语言模型。\n\nOpen-source Model\n它体现开放模型路线，方便社区研究和改造。\n\nOpen weights\n开放权重让用户可以下载、部署和检查模型。\n\nMultilingual AI\n它强调多语言覆盖，服务更广泛的语言场景。", "relations": { "llm": { "label": "属于…", "note": "它本质上是文本生成大模型。" }, "open-source-model": { "label": "体现…路线", "note": "开放发布方便社区研究和改造。" }, "open-weights": { "label": "发布…", "note": "权重开放后才能下载与部署。" }, "multilingual-ai": { "label": "面向…", "note": "它强调覆盖多语言使用场景。" } } } } }, { "id": "api-dynamic-pricing", "name": "API Dynamic Pricing", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "api" }, { "to": "maas-model-as-a-service" }, { "to": "ai-finops" }, { "to": "ai-unit-economics-ai" } ], "track": "today", "i18n": { "en": { "fullName": "AI API Dynamic Pricing", "factExplain": "A way to change API prices as demand, load, or compute cost changes.", "humanExplain": "API dynamic pricing is Uber surge for software. When the AI highway jams, the meter gets jumpy.\n\nIt changes API prices as demand, load, or compute cost moves. You meet it on model platforms and pricing plans.", "humanExplainDisplay": "API dynamic pricing is ==Uber surge== for software.\nWhen the ==AI highway jams==,\nthe meter gets jumpy.\n\nIt changes API prices\nas demand, load, or compute cost moves.\nYou meet it on model platforms\nand pricing plans.", "relationsNarrative": "API\nDynamic pricing changes what developers pay to call an API.\n\nMaaS\nMaaS platforms use it to price model power and busy times.\n\nAI FinOps\nPrices move, so AI FinOps needs tighter tracking and budgets.\n\nAI Unit Economics\nThe price of each call directly affects profit per service.", "relations": { "api": { "label": "changes … call prices", "note": "Dynamic pricing changes the cost of calling an API." }, "maas-model-as-a-service": { "label": "supports … pricing", "note": "MaaS often charges by model power and usage." }, "ai-finops": { "label": "makes … harder", "note": "Moving prices make cost forecasts harder." }, "ai-unit-economics-ai": { "label": "shapes … math", "note": "The price per call affects profit on each service." } } }, "zh": { "fullName": "AI API 动态定价", "factExplain": "随需求、成本或负载调整 API 价格的机制。", "humanExplain": "API 动态定价就是雨天打车：模型越堵、算力越贵，调用费也跟着跳表。\n\n用于模型平台和套餐，成本随负载、需求起伏。", "humanExplainDisplay": "API 动态定价\n就是==雨天打车==：\n模型越堵、算力越贵，\n调用费也跟着==跳表==。\n\n用于模型平台和套餐，\n成本随负载、需求起伏。", "relationsNarrative": "API\n动态定价直接改变开发者调用 API 的成本。\n\nMaaS\nMaaS 平台常用它区分模型能力和服务时段。\n\nAI FinOps\n价格会波动，FinOps 才更需要监控和预算。\n\nAI Unit Economics\n每次调用的价格，直接影响单笔服务毛利。", "relations": { "api": { "label": "改变…调用价格", "note": "动态定价直接作用在 API 计费入口。" }, "maas-model-as-a-service": { "label": "用于…商业化", "note": "MaaS 常按模型能力和用量定价。" }, "ai-finops": { "label": "增加…管理难度", "note": "价格波动会让成本预测更复杂。" }, "ai-unit-economics-ai": { "label": "影响…核算", "note": "单次调用价格决定毛利空间。" } } } } }, { "id": "api", "name": "API", "layer": "L5", "sublayer": "architecture", "era": "2020", "publishedAt": "2026-05-23T10:10:00Z", "relations": [ { "to": "llm" }, { "to": "inference" }, { "to": "function-call" }, { "to": "mcp" } ], "i18n": { "en": { "fullName": "Application Programming Interface", "factExplain": "A standard doorway for software to share data and actions.", "humanExplain": "An API is a drive-thru window for software. You order fries. You do not climb into the kitchen.\n\nAI uses APIs to reach models and work tools. Developers use them to plug those powers into apps.", "humanExplainDisplay": "An API is a ==drive-thru window== for software.\nYou order fries.\nYou do not ==climb into the kitchen==.\n\nAI uses APIs to reach models and work tools.\nDevelopers use them to plug those powers into apps.", "relationsNarrative": "LLM\nAn API lets LLM power enter real apps as a service.\n\nInference\nInference requests often enter a model service through an API.\n\nFunction-calling\nFunction-call uses a set format to trigger the right API.\n\nMCP\nMCP can bundle many APIs into tools a model can use.", "relations": { "llm": { "label": "calls …", "note": "APIs let LLM power enter real apps as a service." }, "inference": { "label": "serves …", "note": "Inference requests often enter model services through an API." }, "function-call": { "label": "carries …", "note": "Function-call uses a set format to trigger the right API." }, "mcp": { "label": "is bundled by …", "note": "MCP can bundle many APIs into tools a model can use." } } }, "zh": { "fullName": "API 接口", "factExplain": "软件之间互相调用能力和数据的标准入口。", "humanExplain": "API 像外卖取餐码：你按格式报单号，后厨才知道给你哪份菜。\n\n它让应用接入模型、支付、地图等服务，是软件互相办事的通道。", "humanExplainDisplay": "API 像==外卖取餐码==：\n你按格式==报单号==，\n后厨才知道给你哪份菜。\n\n它让应用接入模型、支付、\n地图等服务，是软件\n互相办事的通道。", "relationsNarrative": "LLM\nAPI 让 LLM 能力以服务形式进入真实应用。\n\nInference\nInference 请求通常通过 API 进入模型服务。\n\nFunction-calling\nFunction-call 会按约定格式触发对应 API 功能。\n\nMCP\nMCP 可把多个 API 封装成模型可访问的工具集合。", "relations": { "llm": { "label": "用来调用…" }, "inference": { "label": "对外提供…" }, "function-call": { "label": "承载…" } } } } }, { "id": "apple-intelligence", "name": "Apple AI", "layer": "L5", "sublayer": "product", "era": "2024", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "ai-device-ai" }, { "to": "local-llm" }, { "to": "data-privacy" }, { "to": "multimodal" } ], "track": "history", "i18n": { "en": { "fullName": "Apple Intelligence", "factExplain": "Apple’s personal AI system, built to run first on your own device.", "humanExplain": "Apple Intelligence is like your iPhone got a tiny office manager. You mumble half a plan, and it is already sorting the desk.\n\nYou meet it when you write, use Siri, clean alerts, or fix photos. It tries to do more on your device, so your stuff stays closer.", "humanExplainDisplay": "Apple Intelligence is like your iPhone got\na ==tiny office manager==.\nYou mumble ==half a plan==,\nand it is already sorting the desk.\n\nYou meet it when you write,\nuse Siri, clean alerts, or fix photos.\nIt tries to do more on your device,\nso your stuff stays closer.", "relationsNarrative": "AI device\nApple Intelligence is a common example of AI built into the phone and system.\n\nLocal-LLM\nSome Apple Intelligence tasks try to run on local models first.\n\nData-privacy\nApple Intelligence treats privacy as a key promise.\n\nMultimodal AI\nApple Intelligence needs to understand voice, text, and images.", "relations": { "ai-device-ai": { "label": "is a flagship …", "note": "It is a clear example of AI built into a device and system." }, "local-llm": { "label": "runs partly on …", "note": "Some tasks try to use models on the device first." }, "data-privacy": { "label": "puts focus on …", "note": "Privacy is one of its main selling points." }, "multimodal": { "label": "mixes …", "note": "It needs text, voice, and images to work across the system." } } }, "zh": { "fullName": "Apple Intelligence（苹果智能）", "factExplain": "苹果推出的端侧优先个人智能系统。", "humanExplain": "像把 iPhone 练成你办公室老搭子：你话还没说满，它已经懂意思，很多活先在本机悄悄办了。\n\n用于写作、通知整理、图片处理和 Siri 增强，主打更私密、更贴设备。", "humanExplainDisplay": "像把 iPhone 练成你办公室==老搭子==：\n你话还没说满，\n它已经懂意思，\n很多活先在本机==悄悄办了==。\n\n用于写作、通知整理、\n图片处理和 Siri 增强，\n主打更私密、更贴设备。", "relationsNarrative": "AI device\n它是设备端 AI 的代表形态，把能力直接做进手机和系统。\n\nLocal-LLM\n它有些能力尽量在本地模型上完成，减少把数据外发。\n\nData-privacy\n它主打很多请求先本机处理，隐私保护是核心卖点。\n\nMultimodal AI\n它需要同时理解语音、文字和图像，才能融入系统交互。", "relations": { "ai-device-ai": { "label": "属于…代表产品", "note": "它是设备端 AI 落地的典型形态。" }, "local-llm": { "label": "依赖…本地运行", "note": "部分能力尽量在本机模型上完成。" }, "data-privacy": { "label": "强调…保护", "note": "它把隐私当核心卖点之一。" }, "multimodal": { "label": "整合…输入输出", "note": "要同时处理文字、语音和图像。" } } } } }, { "id": "apriori", "name": "Apriori", "layer": "L2", "era": "1994", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "unsupervised-learning" }, { "to": "recommender-system" }, { "to": "big-data" } ], "track": "history", "i18n": { "en": { "fullName": "Apriori Algorithm", "factExplain": "An algorithm that finds common item groups and rules in data.", "humanExplain": "Apriori is like a grocery cashier with a very nosy notebook. It notices chips and salsa keep riding home together.\n\nIt scans old baskets to find common pairings. Stores use it for shopping analysis and simple recommendations.", "humanExplainDisplay": "Apriori is like a ==grocery cashier==\nwith a very nosy notebook.\nIt notices ==chips and salsa==\nkeep riding home together.\n\nIt scans old baskets\nto find common pairings.\nStores use it for shopping analysis\nand simple recommendations.", "relationsNarrative": "Unsupervised Learning\nApriori finds common patterns without labels.\n\nRecommender System\nApriori helps find items people often buy together.\n\nBig Data\nApriori mines rules from large transaction records.", "relations": { "unsupervised-learning": { "label": "is a … method", "note": "It finds patterns without answer labels." }, "recommender-system": { "label": "helps … find pairings", "note": "It can spot items people often buy together." }, "big-data": { "label": "mines rules from …", "note": "It was used early on for large transaction logs." } } }, "zh": { "fullName": "先验算法", "factExplain": "一种挖掘频繁项集和关联规则的算法。", "humanExplain": "Apriori像煎饼摊老板记搭配：加蛋的常配脆饼，下次就顺嘴推荐。\n\n用于购物篮分析和推荐，从记录里捞常见组合。", "humanExplainDisplay": "Apriori像\n煎饼摊老板记搭配：\n==加蛋的常配脆饼==，\n下次就顺嘴推荐。\n\n用于购物篮分析和推荐，\n从记录里\n捞常见组合。", "relationsNarrative": "Unsupervised Learning\nApriori 不靠标签，直接从数据里找常见模式。\n\nRecommender System\nApriori 能发现商品搭配，辅助推荐规则。\n\nBig Data\nApriori 常用于从大量交易记录中挖关联。", "relations": { "unsupervised-learning": { "label": "属于…方法", "note": "不用标签，也能从数据里找规律。" }, "recommender-system": { "label": "支持…发现搭配", "note": "常用来找商品之间的购买关联。" }, "big-data": { "label": "从…中挖规则", "note": "早期数据挖掘常靠它处理交易记录。" } } } } }, { "id": "assisted-driving-ai", "name": "Driving AI", "layer": "L5", "sublayer": "product", "era": "2010s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "embodied-ai" }, { "to": "human-in-the-loop" }, { "to": "robotaxi" } ], "track": "today", "i18n": { "en": { "fullName": "Driver Assistance AI", "factExplain": "AI that watches the road and helps control a car.", "humanExplain": "Driving AI is like a back-seat driver with actual skills. It spots the curb before your bumper gets too friendly.\n\nIt helps your car follow traffic or hold the lane. It can park too, but you must stay ready to take over.", "humanExplainDisplay": "Driving AI is like a ==back-seat driver==\nwith ==actual skills==.\nIt spots the curb\nbefore your bumper gets too friendly.\n\nIt helps your car follow traffic\nor hold the lane.\nIt can park too,\nbut you must stay ready\nto take over.", "relationsNarrative": "Computer Vision\nDriving AI uses Computer Vision to spot lanes, cars, and people.\n\nEmbodied AI\nDriving AI is Embodied AI inside a real car.\n\nHuman-in-the-loop\nDriving AI needs a human to watch and take over.\n\nRobotaxi\nDriving AI is often a step before Robotaxi.", "relations": { "computer-vision": { "label": "sees the road with …", "note": "Cameras help it spot lanes, people, and obstacles." }, "embodied-ai": { "label": "is a real-world use of …", "note": "It lets AI help move a real car." }, "human-in-the-loop": { "label": "needs … as backup", "note": "A person must be ready to take over." }, "robotaxi": { "label": "comes before …", "note": "It is a step before fully driverless rides." } } }, "zh": { "fullName": "辅助驾驶 AI", "factExplain": "用于感知路况并辅助车辆控制的 AI。", "humanExplain": "像副驾坐了个老练代驾，手不抢你方向盘，却会一直盯着路，眼看要蹭了就赶紧提醒你一把。\n\n多用于跟车、车道保持和泊车，能减轻驾驶压力，但仍要司机随时接管。", "humanExplainDisplay": "像副驾坐了个\n老练==代驾==，\n手不抢你方向盘，\n却会一直盯着路，\n==眼看要蹭了==就\n赶紧提醒你一把。\n\n多用于跟车、\n车道保持和泊车，\n能减轻驾驶压力，\n但仍要司机随时接管。", "relationsNarrative": "Computer Vision\n它靠计算机视觉识别车道、车辆和行人等路况。\n\nEmbodied AI\n它属于具身智能在真实车辆上的典型应用场景。\n\nHuman-in-the-loop\n辅助驾驶需要人类持续监管，并在必要时接管。\n\nRobotaxi\n它常被视为通往无人出租车的过渡阶段。", "relations": { "computer-vision": { "label": "依赖…看路", "note": "摄像头识别车道、行人和障碍物。" }, "embodied-ai": { "label": "属于…落地场景", "note": "它让 AI 直接参与现实世界动作。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "关键决策仍需人类随时接管。" }, "robotaxi": { "label": "是…前一步", "note": "辅助驾驶通常比完全无人更保守。" } } } } }, { "id": "atari-learning-environment", "name": "ALE", "layer": "L5", "sublayer": "product", "era": "2013", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "deep-q-network" }, { "to": "openai-gym" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Atari Learning Environment", "factExplain": "A testbed of Atari games for measuring how well AI agents learn.", "humanExplain": "ALE is an old-school arcade exam for AI. The AI only gets a screen and a joystick. Then it learns not to stink.\n\nResearchers use it to test RL. The same Atari games keep the contest fair.", "humanExplainDisplay": "ALE is an ==old-school arcade exam== for AI.\nThe AI only gets ==a screen and a joystick==.\nThen it learns not to stink.\n\nResearchers use it to test RL.\nThe same Atari games keep the contest fair.", "relationsNarrative": "RL\nALE gives RL the same game setup for fair algorithm tests.\n\nDeep Q-Network\nDeep Q-Network used ALE to prove pixels can be enough for game learning.\n\nOpenAI Gym\nOpenAI Gym often wraps ALE in a standard way to run it.\n\nAgent\nAn Agent watches the screen, presses buttons, and learns from the score.", "relations": { "reinforcement-learning": { "label": "tests …", "note": "ALE is a classic test arena for RL." }, "deep-q-network": { "label": "helped prove …", "note": "Deep Q-Network made deep RL famous on ALE games." }, "openai-gym": { "label": "is wrapped by …", "note": "OpenAI Gym often wraps ALE in a standard way to run it." }, "agent": { "label": "challenges …", "note": "An Agent watches the screen, presses buttons, and learns from the score." } } }, "zh": { "fullName": "Atari Learning Environment，Atari 学习环境", "factExplain": "用于评测智能体的 Atari 游戏环境。", "humanExplain": "ALE就像田径场的标准跑道：只给屏幕和手柄，所有AI同场起跑，看谁真练出本事。\n\n它评测强化学习；同一批游戏里，算法公平比身手。", "humanExplainDisplay": "ALE就像==田径标准跑道==：\n只给屏幕和手柄，\n所有AI同场起跑，\n看谁==真练出本事==。\n\n它评测强化学习；\n同一批游戏里，\n算法公平比身手。", "relationsNarrative": "RL\n它为 RL 提供统一游戏环境，方便比较算法。\n\nDeep Q-Network\nDQN 在它上面证明像素输入也能学会玩游戏。\n\nOpenAI Gym\nGym 常把它封装成标准接口，方便调用。\n\nAgent\n智能体在这里看屏幕、按手柄、拿分数反馈。", "relations": { "reinforcement-learning": { "label": "评测…", "note": "它是强化学习经典试炼场。" }, "deep-q-network": { "label": "验证…", "note": "DQN 靠它打响深度强化学习。" }, "openai-gym": { "label": "被…封装", "note": "Gym 常把它封装成可调用环境。" }, "agent": { "label": "考验…", "note": "智能体在游戏里用得分学动作。" } } } } }, { "id": "attention", "name": "Attention", "layer": "L2", "era": "2014", "publishedAt": "2026-05-23T08:55:00Z", "relations": [ { "to": "transformer" }, { "to": "context-window" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Attention Mechanism", "factExplain": "A mechanism for focusing on the most relevant parts during information processing.", "humanExplain": "Attention is the model’s highlighter pen. It marks the juicy bits, not every boring comma.\n\nIt helps the model pick key clues in your prompt. You meet it inside Transformers and LLMs.", "humanExplainDisplay": "Attention is the model’s ==highlighter pen==.\nIt marks the ==juicy bits==,\nnot every boring comma.\n\nIt helps the model pick key clues\nin your prompt.\nYou meet it inside Transformers\nand LLMs.", "relationsNarrative": "Transformer\nAttention helps a Transformer decide which information matters most.\n\nContext-window\nA longer Context-window gives Attention more clues to sort.\n\nLLM\nAttention helps an LLM connect far-apart text.", "relations": { "transformer": { "label": "is key to …", "note": "Transformers use Attention to decide which information matters most." }, "context-window": { "label": "finds key clues in …", "note": "A longer Context-window gives Attention more clues to sort." }, "llm": { "label": "helps … focus", "note": "Attention helps an LLM connect far-apart text." } } }, "zh": { "fullName": "注意力机制", "factExplain": "让模型在处理信息时动态关注更相关部分的机制。", "humanExplain": "注意力机制像刷家庭群：不逐条较真，先盯住谁在说正事、谁在发红包。\n\n它让模型抓住上下文重点，用于翻译、问答和长文本理解。", "humanExplainDisplay": "注意力机制像==刷家庭群==：\n不逐条较真，\n先盯住谁在说正事、\n==谁在发红包==。\n\n它让模型抓住上下文重点，\n用于翻译、问答\n和长文本理解。", "relationsNarrative": "Transformer\nTransformer 依靠 Attention 判断哪些信息更重要。\n\nContext-window\nContext-window 越长，Attention 越需要筛选关键线索。\n\nLLM\nLLM 通过 Attention 建立远距离文本之间的联系。", "relations": { "transformer": { "label": "是…的关键机制" }, "context-window": { "label": "在…中定位重点" }, "llm": { "label": "让…抓住重点" } } } } }, { "id": "autoencoder", "name": "Autoencoder", "layer": "L3", "era": "1986", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "representation-learning" }, { "to": "unsupervised-learning" }, { "to": "variational-autoencoder" }, { "to": "embedding" } ], "track": "history", "i18n": { "en": { "fullName": "Autoencoder", "factExplain": "A neural network that compresses data, then tries to rebuild it.", "humanExplain": "An autoencoder is like stuffing a whole pizza into a tiny lunchbox. Later, it tries to make the pizza look normal again.\n\nPeople use it to shrink data or clean messy data. It also learns compact patterns used in some generative models.", "humanExplainDisplay": "An autoencoder is like stuffing\n==a whole pizza== into ==a tiny lunchbox==.\nLater, it tries to make the pizza\nlook normal again.\n\nPeople use it to shrink data\nor clean messy data.\nIt also learns compact patterns\nused in some generative models.", "relationsNarrative": "Representation Learning\nAn autoencoder learns key representations by compressing data and rebuilding it.\n\nUnsupervised Learning\nAn autoencoder often trains on unlabeled data to find hidden structure.\n\nVAE\nA VAE adds probability modeling to the autoencoder idea.\n\nEmbedding\nAn autoencoder's middle layer can work as a compact vector representation.", "relations": { "representation-learning": { "label": "is a kind of …", "note": "It learns useful representations by compressing and rebuilding data." }, "unsupervised-learning": { "label": "often works in …", "note": "It can learn data structure without human labels." }, "variational-autoencoder": { "label": "develops into …", "note": "A VAE is a probability-based version of an autoencoder." }, "embedding": { "label": "produces …", "note": "Its middle layer can become a compact vector representation." } } }, "zh": { "fullName": "自编码器", "factExplain": "把输入压缩后再重建的神经网络。", "humanExplain": "自编码器像考前偷偷抄小抄：整本书塞成几行关键词，真到上场时，再靠这点线索尽量把原话倒出来。\n\n常用来做压缩、降噪和学特征，也是生成模型常见底座。", "humanExplainDisplay": "自编码器像考前\n偷偷抄==小抄==：\n整本书塞成几行关键词，\n真到上场时，\n再靠这点线索\n尽量把原话\n==倒出来==。\n\n常用来做压缩、降噪\n和学特征，\n也是生成模型常见底座。", "relationsNarrative": "Representation Learning\n它通过压缩再重建，学习数据里的关键表示。\n\nUnsupervised Learning\n它常在无标签数据上训练，用来发现隐藏结构。\n\nVariational Autoencoder\nVAE 在自编码器基础上，加入了概率建模。\n\nEmbedding\n它的中间层输出，常可作为紧凑向量表示。", "relations": { "representation-learning": { "label": "属于…方法", "note": "它靠压缩重建来学习有用表示。" }, "unsupervised-learning": { "label": "常用于…场景", "note": "不靠人工标签也能学数据结构。" }, "variational-autoencoder": { "label": "发展成…", "note": "VAE 是自编码器的概率化版本。" }, "embedding": { "label": "产出…表示", "note": "中间压缩层常被拿来当向量表示。" } } } } }, { "id": "automated-planning", "name": "Planning", "layer": "L2", "era": "1961", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "long-horizon-task" }, { "to": "a-search" }, { "to": "world-model" } ], "track": "history", "i18n": { "en": { "fullName": "Automated Planning", "factExplain": "A way for a system to make action steps toward a goal.", "humanExplain": "Planning is your taco-night friend. You say “make tacos.” It makes the list and remembers napkins.\n\nIt turns a goal into doable steps. Agents and robots use it for long jobs.", "humanExplainDisplay": "Planning is your ==taco-night friend==.\nYou say ==“make tacos.”==\nIt makes the list\nand remembers napkins.\n\nIt turns a goal into doable steps.\nAgents and robots use it\nfor long jobs.", "relationsNarrative": "Agent\nPlanning helps an Agent split a goal into next steps.\n\nLong-horizon\nLong tasks need planning, not just last-minute guesses.\n\nA* Search\nPlanning can use A* Search to find a better path.\n\nWorld model\nA world model lets planning try the plan inside first.", "relations": { "agent": { "label": "plans steps for …", "note": "Agents use planning to turn goals into action plans." }, "long-horizon-task": { "label": "handles …", "note": "Longer tasks need a plan before action starts." }, "a-search": { "label": "finds paths with …", "note": "A* Search can help planning choose a good path." }, "world-model": { "label": "rehearses with …", "note": "A world model lets planning test steps before acting." } } }, "zh": { "fullName": "自动规划", "factExplain": "让系统自动生成达成目标的行动步骤。", "humanExplain": "这活像婚礼总控：你只管说要办成什么样，它负责排流程、卡时间、避免现场翻车。\n\n常用于智能体、机器人和复杂任务执行，把目标拆成可落地步骤。", "humanExplainDisplay": "这活像婚礼总控：\n你只管说要办成什么样，\n它负责排流程、卡时间，\n避免现场==翻车==。\n\n常用于智能体、机器人\n和复杂任务执行，\n把目标拆成可落地步骤。", "relationsNarrative": "Agent\n它常作为智能体的“大脑导航”，负责拆目标和排步骤。\n\nLong-horizon-task\n任务链条越长，越不能只靠临场发挥，更需要提前规划。\n\nA* Search\n很多规划方法会借助搜索，在候选路径里找更优方案。\n\nWorld-model\n如果系统能先在内部预演环境，规划通常会更稳。", "relations": { "agent": { "label": "给…安排步骤", "note": "智能体常靠它把目标变行动计划。" }, "long-horizon-task": { "label": "处理…任务", "note": "任务越长越复杂，越需要先规划。" }, "a-search": { "label": "可用…找方案", "note": "规划常借搜索寻找可行路径。" }, "world-model": { "label": "依赖…预演", "note": "有环境模型时，规划会更靠谱。" } } } } }, { "id": "automated-theorem-proving", "name": "ATP", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "resolution-principle" }, { "to": "unification" }, { "to": "logic-programming" }, { "to": "ai-math-discovery" } ], "track": "history", "i18n": { "en": { "fullName": "Automated Theorem Proving", "factExplain": "A technique for making programs find proofs for math statements.", "humanExplain": "ATP is a math referee with a whistle. You claim the answer scored, and it checks every step on replay.\n\nIt helps prove math claims and check chips. It also checks programs when mistakes must not slip through.", "humanExplainDisplay": "ATP is a ==math referee== with a whistle.\nYou claim the answer scored,\nand it ==checks every step== on replay.\n\nIt helps prove math claims and check chips.\nIt also checks programs\nwhen mistakes must not slip through.", "relationsNarrative": "Resolution\nResolution is a core reasoning rule in many theorem provers.\n\nUnification\nUnification lines up variables so reasoning rules can fit.\n\nLogic\nLogic wraps proof search into a program that can run.\n\nAI Math Discovery\nAutomated Theorem Proving gives AI Math Discovery a strict final check.", "relations": { "resolution-principle": { "label": "reasons with …", "note": "Resolution is a core reasoning rule in many proof tools." }, "unification": { "label": "matches variables with …", "note": "Unification lines up variables so rules fit formulas." }, "logic-programming": { "label": "helps run …", "note": "Logic turns proof search into a working program." }, "ai-math-discovery": { "label": "checks results for …", "note": "AI Math Discovery still needs strict proofs at the end." } } }, "zh": { "fullName": "自动定理证明", "factExplain": "用程序自动寻找数学命题证明的技术。", "humanExplain": "自动定理证明像武林公证人：你说招式无敌，它把每一步谱子验到死。\n\n用于数学、芯片和程序验证，适合绝不能错的场合。", "humanExplainDisplay": "自动定理证明像\n==武林公证人==：\n你说招式无敌，\n它把每一步谱子验到死。\n\n用于数学、芯片，\n和程序验证，\n适合绝不能错的场合。", "relationsNarrative": "Resolution\n归结原理是许多定理证明器的核心推理规则。\n\nUnification\n合一负责把变量对齐，让推理规则能套上。\n\nLogic Programming\n逻辑编程把证明过程包装成可运行的程序。\n\nAI Math Discovery\n自动证明为数学发现提供可验证的收尾。", "relations": { "resolution-principle": { "label": "用…做推理", "note": "归结是许多证明器的核心规则。" }, "unification": { "label": "靠…对齐变量", "note": "合一让规则能准确套到公式上。" }, "logic-programming": { "label": "支撑…运行", "note": "逻辑编程把证明过程变成程序。" }, "ai-math-discovery": { "label": "为…验算结论", "note": "数学发现需要严格证明来收尾。" } } } } }, { "id": "automatic-differentiation", "name": "Autodiff", "layer": "L2", "era": "1964", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "backpropagation" }, { "to": "parameter" }, { "to": "sgd" }, { "to": "framework" } ], "track": "history", "i18n": { "en": { "fullName": "Automatic Differentiation", "factExplain": "A fast way for computers to find derivatives automatically.", "humanExplain": "Autodiff is like Google Maps for a marble on a hill. It points downhill fast, so the marble stops acting lost.\n\nIt finds gradients during neural network training. Then the optimizer uses them to update parameters.", "humanExplainDisplay": "Autodiff is like ==Google Maps for a marble== on a hill.\nIt points ==downhill== fast,\nso the marble stops acting lost.\n\nIt finds gradients during neural network training.\nThen the optimizer uses them\nto update parameters.", "relationsNarrative": "Backpropagation\nBackprop sends errors backward and uses autodiff to find derivatives fast.\n\nParameter\nAutodiff finds gradients, and those gradients guide each parameter change.\n\nSGD\nSGD needs gradients from autodiff before it can update parameters.\n\nFramework\nModern frameworks usually have autodiff built in.", "relations": { "backpropagation": { "label": "helps … find gradients", "note": "Backprop uses autodiff to get layer derivatives fast." }, "parameter": { "label": "guides … updates", "note": "Its gradients show how each parameter should change." }, "sgd": { "label": "hands gradients to …", "note": "SGD needs gradients before it can move parameters." }, "framework": { "label": "is built into …", "note": "Modern frameworks usually include autodiff by default." } } }, "zh": { "fullName": "Automatic Differentiation｜自动微分", "factExplain": "一种高效自动计算导数的方法。", "humanExplain": "像短视频剪完后，软件顺手给你标出哪一刀卡顿、往前挪几帧更顺，省得你一格格回看。\n\n它负责高效算梯度，是训练神经网络的基础，结果会交给优化器更新参数。", "humanExplainDisplay": "像短视频剪完后，\n软件顺手给你标出\n哪一刀==卡顿==、\n往前挪几帧更==顺==，\n省得你一格格回看。\n\n它负责高效算梯度，\n是训练神经网络的基础，\n结果会交给优化器更新参数。", "relationsNarrative": "Backpropagation\n反向传播把误差往回传，靠它高效算各层导数。\n\nParameter\n它算出的梯度，决定每个参数该怎么调整。\n\nSGD\nSGD 要先拿到梯度，才能更新参数。\n\nFramework\n现代深度学习框架通常内置了它。", "relations": { "backpropagation": { "label": "支撑…计算梯度", "note": "反向传播靠它高效算出各层导数。" }, "parameter": { "label": "指导…更新", "note": "它算出的梯度决定参数怎么调。" }, "sgd": { "label": "把结果交给…", "note": "优化器拿到梯度后才知道往哪走。" }, "framework": { "label": "常由…封装", "note": "现代深度学习框架把它做成标配。" } } } } }, { "id": "automation-job", "name": "Automation-job", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T11:00:00Z", "relations": [ { "to": "agent" }, { "to": "ai-anxiety" }, { "to": "agi" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "Automation and Jobs", "factExplain": "How AI automation changes tasks, job roles, and the way people work.", "humanExplain": "Automation is not a robot eating your whole job like a pizza. It snatches the boring slices first: copy-paste chores.\n\nIt changes what your job is worth. People who steer AI tools gain power. Plain repeat work gets risky.", "humanExplainDisplay": "Automation is not a robot\n==eating your whole job like a pizza==.\nIt snatches the boring slices first:\n==copy-paste chores==.\n\nIt changes what your job is worth.\nPeople who steer AI tools gain power.\nPlain repeat work gets risky.", "relationsNarrative": "Agent\nAgents push automation from single steps into complex task work.\n\nAI-anxiety\nAutomation and Jobs can raise AI-anxiety when real roles feel at risk.\n\nAGI\nAGI makes Automation and Jobs seem like a longer structural change.\n\nAI-regulation\nAI-regulation sets boundaries for the disruption from Automation and Jobs.", "relations": { "agent": { "label": "is amplified by …", "note": "Agents can turn small automation into longer task chains." }, "ai-anxiety": { "label": "can trigger …", "note": "Job automation feels scarier when it touches real work." }, "agi": { "label": "looks bigger with …", "note": "AGI makes people picture deeper job changes over time." }, "ai-regulation": { "label": "needs …", "note": "AI-regulation can set limits for job disruption." } } }, "zh": { "fullName": "自动化与就业", "factExplain": "AI 自动化对岗位任务、职业结构和工作方式的影响。", "humanExplain": "自动化岗位像公司里的复读机活儿：每天照流程点点点，AI 来了最先盯上。\n\n它常见于客服、录入、报表等重复流程，关键是人是否还负责判断。", "humanExplainDisplay": "自动化岗位像公司里的==复读机活儿==：\n每天照流程==点点点==，\nAI 来了最先盯上。\n\n它常见于客服、录入、\n报表等重复流程，\n关键是人是否还负责判断。", "relationsNarrative": "Agent\nAgent 让自动化从单点操作扩展到复杂任务执行。\n\nAI-anxiety\nAutomation-job 越接近现实岗位，AI-anxiety 越容易上升。\n\nAGI\nAGI 使 Automation-job 的影响被想象为更长期的结构变化。\n\nAI-regulation\nAI-regulation 为 Automation-job 带来的冲击设定边界。", "relations": { "agent": { "label": "被…放大" }, "ai-anxiety": { "label": "引发…" }, "ai-regulation": { "label": "需要…介入" } } } } }, { "id": "autonomous-ai-chemist", "name": "AI Chemist", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "ai-for-science" }, { "to": "ai-drug-discovery" }, { "to": "physical-ai-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Autonomous AI Chemist", "factExplain": "An AI system for planning, running, and improving chemistry experiments on its own.", "humanExplain": "It is like a robot baker in a very serious kitchen. When the cookies flop, it changes the recipe.\n\nIt helps make new materials and medicines. In robot labs, it works faster. But safety rules must stay in charge.", "humanExplainDisplay": "It is like a ==robot baker==\nin a very serious kitchen.\nWhen the cookies flop,\nit ==changes the recipe==.\n\nIt helps make new materials and medicines.\nIn robot labs,\nit works faster.\nBut safety rules must stay in charge.", "relationsNarrative": "Agent\nIt uses Agent-style planning to turn a goal into experiments.\n\nAI for Science\nIt is a clear example of AI working inside science.\n\nAI Drug Discovery\nIt can make and test candidate molecules by itself.\n\nPhysical AI\nIt needs robot arms and lab tools to do real experiments.", "relations": { "agent": { "label": "plans experiments with …", "note": "It breaks a research goal into experiment steps." }, "ai-for-science": { "label": "fits into …", "note": "It is a key way AI joins science work." }, "ai-drug-discovery": { "label": "speeds up …", "note": "Automated experiments can test candidate molecules faster." }, "physical-ai-ai": { "label": "does experiments through …", "note": "Robot arms and lab tools turn its plans into real tests." } } }, "zh": { "fullName": "自主 AI 化学家", "factExplain": "能自主设计、执行并迭代化学实验的 AI 系统。", "humanExplain": "自主 AI 化学家像会抓药的老中医：开方、煎药、看疗效，药不对还会改方。\n\n用于材料、药物和自动实验室，提速但必须守安全边界。", "humanExplainDisplay": "自主 AI 化学家像\n==会抓药的老中医==：\n开方、煎药、看疗效，\n药不对还会改方。\n\n用于材料、药物\n和自动实验室，\n提速但必须守安全边界。", "relationsNarrative": "Agent\n它用 Agent 式规划，把研究目标拆成实验。\n\nAI For Science\n它是 AI 进入化学科研的典型形态。\n\nAI Drug Discovery\n它能自动合成与筛选候选分子。\n\nPhysical AI\n它要连接机械臂和仪器，才能动手。", "relations": { "agent": { "label": "用…规划实验", "note": "它把目标拆成可执行实验步骤。" }, "ai-for-science": { "label": "属于…场景", "note": "它是 AI 参与科研的典型形态。" }, "ai-drug-discovery": { "label": "加速…流程", "note": "自动实验可缩短候选分子验证。" }, "physical-ai-ai": { "label": "借…动手实验", "note": "机械臂和仪器让计划落地。" } } } } }, { "id": "autoregressive-model", "name": "Autoregressive Model", "layer": "L2", "era": "1950", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "language-modeling" }, { "to": "transformer" }, { "to": "masked-language-modeling" } ], "track": "history", "i18n": { "en": { "fullName": "Autoregressive Model", "factExplain": "A model that predicts the next token from the text before it.", "humanExplain": "It is like phone autocomplete. Type “Dear Principal,” and it probably will not suggest “pizza dragon.”\n\nIt builds text one token at a time. You meet it in chatbots. You meet it in writing tools. You meet it in code autocomplete.", "humanExplainDisplay": "It is like ==phone autocomplete==.\nType “Dear Principal,”\nand it probably will not suggest ==“pizza dragon.”==\n\nIt builds text one token at a time.\nYou meet it in chatbots.\nYou meet it in writing tools.\nYou meet it in code autocomplete.", "relationsNarrative": "Token\nIt predicts Tokens in order to build the full text.\n\nLM\nLM often teaches it to continue with the next token.\n\nTransformer\nModern autoregressive models often use a Transformer.\n\nMLM\nIt writes forward, while MLM fills in blanks.", "relations": { "token": { "label": "predicts each …", "note": "It adds one Token at a time to continue the text." }, "language-modeling": { "label": "is used for …", "note": "LM often trains a model to predict the next token." }, "transformer": { "label": "is often built with …", "note": "Many modern autoregressive models use a Transformer." }, "masked-language-modeling": { "label": "contrasts with …", "note": "It writes forward, while MLM fills in blanks." } } }, "zh": { "fullName": "自回归模型", "factExplain": "按前文条件，逐步预测下一个 token 的生成模型。", "humanExplain": "像接龙讲鬼故事，前一句刚吓完，后一句只能顺着往下编；开头挖了坑，后面都得接着填。\n\n它按前文一步步生成，常用于对话、写作和代码补全。", "humanExplainDisplay": "像接龙讲鬼故事，\n前一句刚吓完，\n后一句只能顺着往下编；\n开头挖了==坑==，\n后面都得接着==填==。\n\n它按前文一步步生成，\n常用于对话、\n写作和代码补全。", "relationsNarrative": "Token\n它按顺序逐个预测 token，拼出完整文本。\n\nLanguage Modeling\n语言建模的经典目标，就是让它学会续写下一个 token。\n\nTransformer\n现代自回归模型通常用 Transformer 来提升长序列建模能力。\n\nMasked-language-modeling\n它强调顺着往后生成，MLM 更像先挖空再填空。", "relations": { "token": { "label": "逐个预测…", "note": "它一次生成一个 token 往后接。" }, "language-modeling": { "label": "常用于…任务", "note": "语言建模常让它预测下一个词元。" }, "transformer": { "label": "常由…实现", "note": "现代这类模型多用 Transformer 搭建。" }, "masked-language-modeling": { "label": "对比…范式", "note": "一个按顺序生成，一个靠遮盖补空。" } } } } }, { "id": "background-agent", "name": "Background Agent", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "long-horizon-task" }, { "to": "agentic-coding" }, { "to": "agent-harness" } ], "track": "today", "i18n": { "en": { "fullName": "Background AI Agent", "factExplain": "An AI agent that runs tasks in the background while you do other things.", "humanExplain": "A Background Agent is like a slow cooker with a to-do list. You leave, and it keeps bubbling away.\n\nIt helps with code fixes and research cleanup. Keep its access tight, and check its work.", "humanExplainDisplay": "A Background Agent is like a ==slow cooker==\nwith a to-do list.\nYou leave,\nand it ==keeps bubbling away==.\n\nIt helps with code fixes\nand research cleanup.\nKeep its access tight,\nand check its work.", "relationsNarrative": "Agent\nA Background Agent is an Agent that keeps working after you leave.\n\nLong-horizon\nLong-horizon tasks are a good fit because they take time.\n\nAgentic coding\nAgentic coding often uses Background Agents for cloud code fixes.\n\nAgent harness\nAn Agent harness manages its access, logs, and state.", "relations": { "agent": { "label": "builds on …", "note": "It is an Agent that keeps working after you leave." }, "long-horizon-task": { "label": "works on …", "note": "Long tasks can run in the background without making you wait." }, "agentic-coding": { "label": "is used for …", "note": "Cloud code fixes often use Background Agents." }, "agent-harness": { "label": "runs inside …", "note": "An Agent harness manages access, logs, and job state." } } }, "zh": { "fullName": "后台 AI 代理", "factExplain": "一种在后台异步执行任务的 AI 代理。", "humanExplain": "后台代理像把任务塞进自助洗衣机：人走去睡觉，它还在后台转到完工。\n\n适合代码修复、资料整理等长任务，需控权限和验收。", "humanExplainDisplay": "后台代理像把任务\n塞进==自助洗衣机==：\n人走去睡觉，\n它还在后台==转到完工==。\n\n适合代码修复、资料整理等长任务，\n需控权限和验收。", "relationsNarrative": "Agent\n后台代理是 Agent 的异步工作形态。\n\nLong-horizon\n长周期任务适合交给后台代理持续推进。\n\nAgentic coding\n后台代理常用于云端代码修改和修复。\n\nAgent harness\n运行环境负责托管权限、日志和状态。", "relations": { "agent": { "label": "以…为基础", "note": "后台代理是 Agent 的异步形态。" }, "long-horizon-task": { "label": "处理…", "note": "长任务适合交给后台慢慢跑。" }, "agentic-coding": { "label": "常用于…", "note": "云端改代码常用这种模式。" }, "agent-harness": { "label": "依赖…运行", "note": "运行、权限和日志都要托管。" } } } } }, { "id": "backpropagation-through-time", "name": "BPTT", "layer": "L2", "era": "1990", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "backpropagation" }, { "to": "sequence-modeling" }, { "to": "lstm" }, { "to": "automatic-differentiation" } ], "track": "history", "i18n": { "en": { "fullName": "Backpropagation Through Time", "factExplain": "A way to train sequence models by unrolling time and sending errors backward.", "humanExplain": "BPTT is like replaying a messy group chat after a blowup. The last emoji gets blamed, but the drama started ten messages ago.\n\nIt trains AI for ordered data, like text or speech. It traces a later mistake back to earlier steps.", "humanExplainDisplay": "BPTT is like replaying a ==messy group chat== after a blowup.\nThe last emoji gets blamed,\nbut the drama started ==ten messages ago==.\n\nIt trains AI for ordered data,\nlike text or speech.\nIt traces a later mistake\nback to earlier steps.", "relationsNarrative": "Backpropagation\nBPTT is Backpropagation extended across time steps.\n\nSequence Modeling\nSequence Modeling uses BPTT to send later errors back to earlier steps.\n\nLSTM\nLSTM networks often use BPTT during training.\n\nautodiff\nAutodiff often handles the gradient math for BPTT.", "relations": { "backpropagation": { "label": "is the time version of …", "note": "BPTT extends Backpropagation across unfolded time steps." }, "sequence-modeling": { "label": "trains …", "note": "Sequence tasks use it to push errors back through earlier steps." }, "lstm": { "label": "often trains …", "note": "LSTM networks often learn with BPTT." }, "automatic-differentiation": { "label": "is often done by …", "note": "Modern tools often use autodiff to compute these gradients." } } }, "zh": { "fullName": "时间反向传播（Backpropagation Through Time）", "factExplain": "把序列模型在时间维展开后做反向传播的训练方法。", "humanExplain": "吵架吵崩了，不能只怪最后一句嘴快，前面每次拱火都得倒回去算，这才知道锅从哪来。\n\n常用于训练文本、语音等序列模型，把后面的误差一路追到前面。", "humanExplainDisplay": "吵架吵崩了，\n不能只怪\n==最后一句嘴快==，\n前面每次拱火\n都得==倒回去算==，\n这才知道锅从哪来。\n\n常用于训练文本、语音等\n序列模型，\n把后面的误差一路\n追到前面。", "relationsNarrative": "Backpropagation\n它是反向传播在序列与时间展开场景下的延伸。\n\nSequence Modeling\n序列建模常用它把后面的误差传回前面步骤。\n\nLSTM\nLSTM 这类循环网络通常靠它完成训练。\n\nAutodiff\n现代框架常借助自动求导实现这套求梯度过程。", "relations": { "backpropagation": { "label": "是…的时序版", "note": "把普通反传扩展到时间展开的序列。" }, "sequence-modeling": { "label": "用于训练…", "note": "序列任务常靠它把误差往前追。" }, "lstm": { "label": "常用于训练…", "note": "LSTM 这类循环结构常用它学习。" }, "automatic-differentiation": { "label": "常由…实现", "note": "现代框架通常靠自动求导完成计算。" } } } } }, { "id": "backpropagation", "name": "Backpropagation", "layer": "L1", "era": "1986", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "sgd" }, { "to": "deep-learning" }, { "to": "parameter" } ], "track": "history", "i18n": { "en": { "fullName": "Backpropagation", "factExplain": "A training method that sends errors backward to update a model’s parameters.", "humanExplain": "Backprop is a coach with instant replay. After a missed shot, it fixes the wrist, then the elbow.\n\nIt tells a neural network which parts caused the mistake. It is a core step for training deep models.", "humanExplainDisplay": "Backprop is a ==coach with instant replay==.\nAfter a ==missed shot==,\nit fixes the wrist,\nthen the elbow.\n\nIt tells a neural network\nwhich parts caused the mistake.\nIt is a core step\nfor training deep models.", "relationsNarrative": "Neural-network\nBackprop sends the error backward through the Neural-network and helps change its weights.\n\nSGD\nBackprop finds which way to change, then SGD updates the parameters.\n\nDeep Learning\nBackprop makes many-layer training possible in Deep Learning.\n\nParameter\nBackprop tells each Parameter how it should change.", "relations": { "neural-network": { "label": "trains … parameters", "note": "It shows the neural network which weights need changing." }, "sgd": { "label": "works with … to update", "note": "It finds which way to change, then SGD moves the parameters." }, "deep-learning": { "label": "powers … training", "note": "Deep Learning uses backprop to train many-layer networks." }, "parameter": { "label": "decides how … changes", "note": "The backward error tells each parameter how to change." } } }, "zh": { "fullName": "Backpropagation（反向传播）", "factExplain": "一种把输出误差逐层传回去更新参数的训练方法。", "humanExplain": "像师傅带徒弟练刀：不是只说砍歪了，而是从手腕到步子，一路往回纠动作。\n\n它负责告诉网络哪里该改，是训练深度模型的核心步骤。", "humanExplainDisplay": "像师傅带徒弟练刀：\n不是只说==砍歪了==，\n而是从手腕到步子，\n一路往回==纠动作==。\n\n它负责告诉网络哪里该改，\n是训练深度模型的核心步骤。", "relationsNarrative": "Neural-network\n它让神经网络把错误逐层往回传，并调整权重。\n\nSgd\n它先算出该往哪改，再由 SGD 按梯度更新参数。\n\nDeep Learning\n深度学习能训练很多层网络，离不开反向传播。\n\nParameter\n反向传播的核心作用，就是告诉参数该怎么改。", "relations": { "neural-network": { "label": "训练…参数", "note": "它让神经网络知道该改哪些权重。" }, "sgd": { "label": "配合…更新", "note": "算出梯度后，常交给 SGD 真正改参数。" }, "deep-learning": { "label": "支撑…训练", "note": "现代深度学习基本靠它高效训练。" }, "parameter": { "label": "决定…怎么改", "note": "反向传回的误差会指导参数更新。" } } } } }, { "id": "bag-of-words", "name": "Bag-of-Words", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "bm25" }, { "to": "word2vec" }, { "to": "embedding" } ], "track": "history", "i18n": { "en": { "fullName": "Bag-of-Words", "factExplain": "A classic way to turn text into a vector of word counts.", "humanExplain": "Bag-of-Words treats an essay like a Halloween candy bag. It counts each candy type, not the order you grabbed them.\n\nPeople use it to sort texts and search them. It starts fast, but misses word order and context.", "humanExplainDisplay": "Bag-of-Words treats an essay like a ==Halloween candy bag==.\nIt ==counts each candy type==,\nnot the order you grabbed them.\n\nPeople use it to sort texts\nand search them.\nIt starts fast,\nbut misses word order and context.", "relationsNarrative": "Token\nBag-of-Words splits text into tokens, then counts how often they appear.\n\nBM25\nBM25 keeps the word-count idea and improves it for search.\n\nWord2Vec\nWord2Vec helps fix Bag-of-Words because it can catch similar meanings.\n\nEmbedding\nEmbeddings pack meaning better than simple word counts.", "relations": { "token": { "label": "splits into … first", "note": "The text becomes tokens before Bag-of-Words counts them." }, "bm25": { "label": "forms a base for …", "note": "BM25 builds on the same idea of counting word terms." }, "word2vec": { "label": "was partly replaced by …", "note": "Word2Vec can catch similar meanings, not just count words." }, "embedding": { "label": "can upgrade into …", "note": "Embeddings move from sparse counts to dense meaning vectors." } } }, "zh": { "fullName": "词袋模型", "factExplain": "把文本表示成词频向量的经典方法。", "humanExplain": "它看文章，像班长收班费：只记谁交了几次、交多少钱，不管发言顺序，主打一个只认数。\n\n常用于分类、检索等文本任务；上手快，但不懂语序和上下文。", "humanExplainDisplay": "它看文章，\n像班长收==班费==：\n只记谁交了几次、\n交多少钱，\n不管发言顺序，\n主打一个==只认数==。\n\n常用于分类、\n检索等文本任务；\n上手快，但不懂\n语序和上下文。", "relationsNarrative": "Token\n词袋先把文本拆成词或词元，再统计出现次数。\n\nBM25\nBM25 延续词项统计思想，是检索里的经典改进版。\n\nWord2Vec\n词向量弥补了词袋不懂语义相近的老毛病。\n\nEmbedding\n嵌入表示比词袋更能压缩语义信息。", "relations": { "token": { "label": "先拆成…统计", "note": "先把文本切成词或词元再计数。" }, "bm25": { "label": "是…的基础表示", "note": "BM25 建在词项统计这套思路上。" }, "word2vec": { "label": "被…部分替代", "note": "后者能表示语义相近，不只会数词。" }, "embedding": { "label": "可升级为…表示", "note": "从稀疏计数走向稠密语义向量。" } } } } }, { "id": "bahdanau-attention", "name": "Bahdanau Attention", "layer": "L3", "era": "2014", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "attention" }, { "to": "seq2seq" }, { "to": "self-attention" }, { "to": "transformer" } ], "track": "history", "i18n": { "en": { "fullName": "Bahdanau Attention", "factExplain": "An attention method for lining up generated words with input positions.", "humanExplain": "Bahdanau Attention is like a kid doing homework with one finger on the textbook. For each answer, the finger scoots to the right line.\n\nYou meet it in older translation models. As each word is written, it looks back at the most useful input word.", "humanExplainDisplay": "Bahdanau Attention is like\na kid doing homework with ==one finger on the textbook==.\nFor each answer,\nthe finger ==scoots to the right line==.\n\nYou meet it in older translation models.\nAs each word is written,\nit looks back at the most useful input word.", "relationsNarrative": "Attention\nBahdanau Attention is an early classic form of Attention.\n\nSeq2Seq\nIt gives Seq2Seq a moving link between input and output.\n\nSelf-Attention\nSelf-Attention extends this focus idea inside the sequence.\n\nTransformer\nTransformer keeps the attention idea and makes it much bigger.", "relations": { "attention": { "label": "is an early form of …", "note": "Bahdanau Attention is a classic early attention method." }, "seq2seq": { "label": "adds alignment to …", "note": "It gives encoder-decoder models a moving focus." }, "self-attention": { "label": "led toward …", "note": "Self-Attention applies the focus idea inside one sequence." }, "transformer": { "label": "helped inspire …", "note": "It helped pave the road for attention-based architectures." } } }, "zh": { "fullName": "Bahdanau Attention（Bahdanau 注意力）", "factExplain": "一种在生成时动态对齐输入位置的注意力机制。", "humanExplain": "它像开卷考试答题，写到哪题就翻哪页，不用整本硬背，现用现找最对口那段。\n\n常见于翻译等序列生成，让输出时动态对齐输入重点。", "humanExplainDisplay": "它像==开卷考试==答题，\n写到哪题就翻哪页，\n不用整本硬背，\n现用现找最==对口==那段。\n\n常见于翻译等序列生成，\n让输出时动态对齐输入重点。", "relationsNarrative": "Attention\n它是经典注意力机制的早期代表形式。\n\nSeq2Seq\n它为 Seq2Seq 补上输入输出的动态对齐能力。\n\nSelf-Attention\nSelf-Attention 把这种选重点思路扩展到序列内部。\n\nTransformer\nTransformer 延续并放大了注意力这条路线。", "relations": { "attention": { "label": "属于…早期形式", "note": "它是经典注意力机制代表之一" }, "seq2seq": { "label": "增强…对齐", "note": "为编码器解码器补上动态对焦" }, "self-attention": { "label": "被…进一步发展", "note": "后者把注意力扩展到序列内部" }, "transformer": { "label": "启发…设计", "note": "为后来的注意力架构铺路" } } } } }, { "id": "base-model", "name": "Base model", "layer": "L3", "era": "2022", "publishedAt": "2026-05-29T16:08:01.211Z", "relations": [ { "to": "foundation-model" }, { "to": "pretraining" }, { "to": "fine-tuning" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Base model", "factExplain": "A general AI model pretrained on huge data and later adapted for many tasks.", "humanExplain": "A base model is a frozen cheese pizza. The hard part is done, but it still needs toppings before anyone cheers.\n\nIt sits under many AI products. After fine-tuning, it can become a helper for one job.", "humanExplainDisplay": "A base model is a ==frozen cheese pizza==.\nThe ==hard part is done==,\nbut it still needs toppings\nbefore anyone cheers.\n\nIt sits under many AI products.\nAfter fine-tuning,\nit can become a helper for one job.", "relationsNarrative": "Foundation-model\nBase model is close to Foundation-model. Both mean a general AI base.\n\nPretraining\nPretraining gives a Base model its basic skills.\n\nFine-tuning\nFine-tuning shapes a Base model for one task.\n\nLLM\nMany LLMs are Base models focused on language.", "relations": { "foundation-model": { "label": "is close to …", "note": "Both mean a general AI base model." }, "pretraining": { "label": "starts with …", "note": "Pretraining gives the Base model its basic skills." }, "fine-tuning": { "label": "gets specialized by …", "note": "Fine-tuning makes it fit a specific task or setting." }, "llm": { "label": "can become …", "note": "Many LLMs are text-focused Base models." } } }, "zh": { "fullName": "基础模型", "factExplain": "经大规模预训练、可继续适配多任务的通用模型。", "humanExplain": "说白了，它是块已经打好的毛坯房：水电框架都在，但要住成办公室还是咖啡馆，还得继续装修。\n\n它是很多 AI 产品的底座，常经微调后变成特定场景助手。", "humanExplainDisplay": "说白了，它是块已经打好的==毛坯房==：\n水电框架都在，\n但要住成办公室还是咖啡馆，\n还得继续==装修==。\n\n它是很多 AI 产品的底座，\n常经微调后变成特定场景助手。", "relationsNarrative": "Foundation-model\nBase model 常被视作 Foundation-model 的近义说法，强调“通用底座”这层意思。\n\nPretraining\nPretraining 决定 Base model 的基础能力，是它成型的关键阶段。\n\nFine-tuning\nFine-tuning 会在 Base model 之上继续雕刻，让它更贴合具体任务。\n\nLLM\n很多 LLM 都是 Base model 的一种，聚焦语言理解与生成。", "relations": { "foundation-model": { "label": "常被视作…", "note": "两者语义很近，都是通用底座模型。" }, "pretraining": { "label": "由…打底", "note": "基础能力主要来自大规模预训练阶段。" }, "fine-tuning": { "label": "经…变专长", "note": "微调让它更适合具体任务和场景。" }, "llm": { "label": "可发展成…", "note": "很多 LLM 本质上是文本类基础模型。" } } } } }, { "id": "batch-normalization", "name": "Batch Normalization", "layer": "L2", "era": "2015", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "layer-normalization" }, { "to": "gradient-descent" }, { "to": "cnn" }, { "to": "resnet" } ], "track": "history", "i18n": { "en": { "fullName": "Batch Normalization", "factExplain": "A technique that normalizes neuron signals in each training batch.", "humanExplain": "Batch Normalization is like a factory resetting every part to the same baseline. No bolt starts too long or too short before assembly.\n\nIt evens out neuron signals in each training batch. Deep nets use it to learn faster and wobble less.", "humanExplainDisplay": "Batch Normalization is like a factory\nresetting every part\nto the ==same baseline==.\nNo bolt starts too long\nor too short before ==assembly==.\n\nIt evens out neuron signals\nin each training batch.\nDeep nets use it\nto learn faster\nand wobble less.", "relationsNarrative": "Layer Normalization\nBoth steady training, but Batch Normalization uses batch stats while Layer Normalization works per sample.\n\nGradient Descent\nBatch Normalization can reduce training wobble, so Gradient Descent can settle more easily.\n\nCNN\nBatch Normalization was widely used in CNNs to make training faster and steadier.\n\nResNet\nBatch Normalization often appears inside deep ResNet models.", "relations": { "layer-normalization": { "label": "is often compared with …", "note": "Both normalize values, but they use different groups." }, "gradient-descent": { "label": "helps … settle", "note": "It makes gradient updates a bit steadier." }, "cnn": { "label": "is common in … training", "note": "It was once a standard part of many CNNs." }, "resnet": { "label": "often appears in …", "note": "Many ResNet designs use it." } } }, "zh": { "fullName": "批量归一化", "factExplain": "在训练中按批次标准化激活值的技术。", "humanExplain": "批量归一化像工厂出货前统一校准：先把整批零件拉到同一基准线，后面组装时才不会东歪西扭。\n\n常用于训练深层网络，帮助收敛更稳更快。", "humanExplainDisplay": "批量归一化像工厂出货前\n统一==校准==：\n先把整批零件拉到同一基准线，\n后面组装时才不至于==东歪西扭==。\n\n常用于训练深层网络，\n帮助收敛更稳更快。", "relationsNarrative": "Layer Normalization\n两者都在稳训练，但一个看批次，一个看单样本。\n\nGradient Descent\n它能缓和训练波动，让梯度下降更容易收敛。\n\nCNN\n它曾广泛用于卷积网络，加速并稳定训练过程。\n\nResNet\nResNet 这类深层网络里，经常能看到它的身影。", "relations": { "layer-normalization": { "label": "常与…对比", "note": "两者都做归一化，但统计方式不同。" }, "gradient-descent": { "label": "辅助…收敛", "note": "它能让梯度更新更稳定一些。" }, "cnn": { "label": "常用于…训练", "note": "它曾是卷积网络里的常见标配。" }, "resnet": { "label": "常见于…结构", "note": "很多残差网络都会配合它使用。" } } } } }, { "id": "bayesian-network", "name": "Bayesian Network", "layer": "L2", "era": "1985", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "belief-state" }, { "to": "knowledge-graph" }, { "to": "structural-causal-model" }, { "to": "do-calculus" } ], "track": "history", "i18n": { "en": { "fullName": "Bayesian Network", "factExplain": "A probability model that uses arrows to show how uncertain things depend on each other.", "humanExplain": "A Bayesian Network is a detective’s corkboard for everyday chaos. Burnt toast has arrows to Dad and one guilty toaster.\n\nIt draws arrows between uncertain things. You meet it in diagnosis and risk checks. It also helps with cause-and-effect questions.", "humanExplainDisplay": "A Bayesian Network is a ==detective’s corkboard==\nfor everyday chaos.\nBurnt toast has ==arrows==\nto Dad and one guilty toaster.\n\nIt draws arrows between uncertain things.\nYou meet it in diagnosis and risk checks.\nIt also helps with cause-and-effect questions.", "relationsNarrative": "Belief State\nA Bayesian Network can show a system's uncertain beliefs as probabilities.\n\nKnowledge Graph\nBoth use graphs, but a Bayesian Network focuses on probability links.\n\nSCM\nBoth show links between variables, but an SCM focuses on causes.\n\nDo-Calculus\nWhen the graph asks about interventions, it can connect to Do-Calculus.", "relations": { "belief-state": { "label": "represents …", "note": "It can store unsure states as probabilities in a graph." }, "knowledge-graph": { "label": "uses graphs like …", "note": "A Bayesian Network tracks probability links; a Knowledge Graph tracks facts." }, "structural-causal-model": { "label": "shares structure with …", "note": "Both map variable links, but an SCM focuses on causes." }, "do-calculus": { "label": "can connect to …", "note": "With intervention meaning, the graph moves toward causal analysis." } } }, "zh": { "fullName": "贝叶斯网络", "factExplain": "用有向图表示随机变量条件依赖关系的概率模型。", "humanExplain": "贝叶斯网络像家族群里理亲戚账：谁会影响谁、消息先传到哪，一根根线顺下来就不乱了。\n\n常用于诊断、风险判断和因果推理，适合处理不确定信息。", "humanExplainDisplay": "贝叶斯网络像\n家族群里理==亲戚账==：\n谁会影响谁、\n消息先传到哪，\n一根根线顺下来\n就==不乱了==。\n\n常用于诊断、\n风险判断和因果推理，\n适合处理不确定信息。", "relationsNarrative": "Belief State\n它常被用来表示系统对未知状态的概率性判断。\n\nKnowledge Graph\n两者都用图表达关系，但它更强调概率依赖。\n\nStructural Causal Model\n两者都描述变量结构，但后者更强调因果机制。\n\nDo-Calculus\n当图里涉及干预问题时，会进一步连到它。", "relations": { "belief-state": { "label": "常用于表示…", "note": "可把不确定状态编码成概率结构。" }, "knowledge-graph": { "label": "都用图表达", "note": "一个偏概率依赖，一个偏事实关系。" }, "structural-causal-model": { "label": "启发…建模", "note": "两者都关心变量之间的结构关系。" }, "do-calculus": { "label": "可连接到…", "note": "加入干预语义后更接近因果分析。" } } } } }, { "id": "bayesian-optimization", "name": "Bayesian Optimization", "layer": "L2", "era": "1998", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "hyperparameter-optimization" }, { "to": "gaussian-process" }, { "to": "exploration-exploitation" }, { "to": "optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Bayesian Optimization", "factExplain": "An optimization method that uses a probability model to pick the next trial.", "humanExplain": "Bayesian Optimization is like finding the best donut in a huge box. You only get a few bites, so each bite points to the next one.\n\nIt is used for HPO and costly lab tests. It finds good results with fewer tries.", "humanExplainDisplay": "Bayesian Optimization is like finding the ==best donut==\nin a huge box.\nYou only get ==a few bites==,\nso each bite points to the next one.\n\nIt is used for HPO\nand costly lab tests.\nIt finds good results\nwith fewer tries.", "relationsNarrative": "HPO\nBayesian Optimization helps HPO find good settings with fewer training runs.\n\nGP\nA GP often acts as the stand-in model for what to try next.\n\nExploration-Exploitation Tradeoff\nBayesian Optimization uses uncertainty to balance trying new spots and using good spots.\n\nOptimization\nBayesian Optimization is an Optimization method for problems where each test costs a lot.", "relations": { "hyperparameter-optimization": { "label": "helps … try less", "note": "HPO treats each training run like an expensive test." }, "gaussian-process": { "label": "often uses … as a stand-in", "note": "A GP draws a probability map of the unknown function." }, "exploration-exploitation": { "label": "balances …", "note": "It tries new spots and uses known good spots." }, "optimization": { "label": "is a kind of …", "note": "The goal is to find the best result with few tests." } } }, "zh": { "fullName": "贝叶斯优化", "factExplain": "用概率模型选择下一次评估的优化方法。", "humanExplain": "贝叶斯优化像抓娃娃先看爪劲：币不多，边试边估，下一把押最有戏的位置。\n\n用于调超参数和昂贵实验，用更少尝试找好结果。", "humanExplainDisplay": "贝叶斯优化像抓娃娃先看爪劲：\n==币不多==，\n边试边估，\n下一把押==最有戏的位置==。\n\n用于调超参数和昂贵实验，\n用更少尝试，\n找好结果。", "relationsNarrative": "Hyperparameter Optimization\n贝叶斯优化常用于超参数搜索，少训几次也能找到好配置。\n\nGaussian Process\n高斯过程常当代理模型，估计哪里值得下一试。\n\nExploration-Exploitation Tradeoff\n它用概率不确定性，在探索和利用之间做取舍。\n\nOptimization\n它是黑盒优化方法，适合单次评估很贵的问题。", "relations": { "hyperparameter-optimization": { "label": "帮…少试几轮", "note": "调参常把每次训练当昂贵实验。" }, "gaussian-process": { "label": "常用…做代理", "note": "GP 给未知函数补上概率地图。" }, "exploration-exploitation": { "label": "平衡…取舍", "note": "既试新区域，也利用已知好点。" }, "optimization": { "label": "属于…方法", "note": "目标是在少量试验中找最优。" } } } } }, { "id": "beam-search", "name": "Beam Search", "layer": "L2", "era": "1970s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "seq2seq" }, { "to": "autoregressive-model" }, { "to": "inference" }, { "to": "token" } ], "track": "history", "i18n": { "en": { "fullName": "Beam Search", "factExplain": "A decoding method that keeps a few top paths at each step.", "humanExplain": "Beam Search is like a GPS trying a few good routes at once. It kicks out bad routes before you end up behind a mall.\n\nAI uses it for translation and speech text. It is steadier than one guess, but costs more computer work.", "humanExplainDisplay": "Beam Search is like a GPS\ntrying ==a few good routes== at once.\nIt kicks out ==bad routes==\nbefore you end up behind a mall.\n\nAI uses it for translation and speech text.\nIt is steadier than one guess,\nbut costs more computer work.", "relationsNarrative": "Seq2Seq\nBeam Search helps Seq2Seq decode better sequences.\n\nAutoregressive Model\nBeam Search keeps several paths as the model writes step by step.\n\nInference\nBeam Search runs during Inference, not during training.\n\nToken\nBeam Search expands one Token at a time and scores the whole path.", "relations": { "seq2seq": { "label": "decodes for …", "note": "Seq2Seq often uses it to make steadier sequences." }, "autoregressive-model": { "label": "picks next words for …", "note": "It keeps several paths during word-by-word generation." }, "inference": { "label": "runs during …", "note": "It happens when the model is making an answer." }, "token": { "label": "expands by …", "note": "Each step compares possible next tokens." } } }, "zh": { "fullName": "束搜索", "factExplain": "一种每步保留少量高分候选的序列解码方法。", "humanExplain": "下棋时它不只盯眼前一步，会先留几条像样的后手继续算，差的立刻砍掉，免得越下越离谱。\n\n常用于机器翻译、语音识别等生成解码，通常更稳，但也更耗算力。", "humanExplainDisplay": "下棋时它不只盯\n眼前一步，\n会先留几条==像样的后手==\n继续算，差的立刻砍掉，\n免得越下越==离谱==。\n\n常用于机器翻译、\n语音识别等生成解码，\n通常更稳，但也更耗算力。", "relationsNarrative": "Seq2Seq\nSeq2Seq 常用它做解码，提升生成质量。\n\nAutoregressive Model\n它在自回归生成中为每步保留多条候选路径。\n\nInference\n它属于推理阶段的输出搜索策略，不参与训练。\n\nToken\n它按词元一步步展开，并比较整句累计分数。", "relations": { "seq2seq": { "label": "常用于…解码", "note": "Seq2Seq 常用它生成更稳的序列。" }, "autoregressive-model": { "label": "给…挑下文", "note": "它在逐词生成时保留多个候选。" }, "inference": { "label": "属于…阶段", "note": "它发生在模型出答案的时候。" }, "token": { "label": "按…逐步扩展", "note": "每一步都在比较下一个词元候选。" } } } } }, { "id": "belief-propagation", "name": "BP", "layer": "L2", "era": "1982", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "bayesian-network" }, { "to": "probabilistic-graphical-model" }, { "to": "variational-inference" }, { "to": "graph-neural-network" } ], "track": "history", "i18n": { "en": { "fullName": "Belief Propagation", "factExplain": "A method that passes local messages in a graph to estimate probabilities.", "humanExplain": "BP is like a homework group chat. Each kid sends one sure answer, and the whole sheet starts making sense.\n\nIn graph models, it passes small probability messages. You meet it in inference and error-correcting codes. On neat graphs, it is fast and accurate.", "humanExplainDisplay": "BP is like a ==homework group chat==.\nEach kid sends ==one sure answer==,\nand the whole sheet starts making sense.\n\nIn graph models,\nit passes small probability messages.\nYou meet it in inference and error-correcting codes.\nOn neat graphs,\nit is fast and accurate.", "relationsNarrative": "Bayesian Network\nBP passes messages in a Bayesian Network to estimate node probabilities.\n\nPGM\nBP is a classic inference algorithm for PGMs.\n\nVI\nBP and VI both estimate probabilities when exact math is too hard.\n\nGNN\nGNN message passing has a family resemblance to BP.", "relations": { "bayesian-network": { "label": "runs inference in …", "note": "BP passes local probabilities in Bayesian Networks." }, "probabilistic-graphical-model": { "label": "is a classic algorithm for …", "note": "BP is a core inference method in PGMs." }, "variational-inference": { "label": "shares approximate inference with …", "note": "Both estimate the answer when exact probability math is too hard." }, "graph-neural-network": { "label": "inspired message passing in …", "note": "GNNs borrow the idea of nodes sending messages." } } }, "zh": { "fullName": "Belief Propagation／置信传播", "factExplain": "一种在图模型中传递局部信息来估计概率分布的方法。", "humanExplain": "它像村里挨家传话报平安：每户只往隔壁报自己那条准信，来回传几轮，全村的消息就对齐了。\n\n常用于图模型推断、纠错解码；图规整时又快又准。", "humanExplainDisplay": "它像村里挨家==传话报平安==：\n每户只往隔壁\n报自己那条准信，\n来回传几轮，\n全村的消息就==对齐了==。\n\n常用于图模型推断、\n纠错解码；\n图规整时又快又准。", "relationsNarrative": "Bayesian Network\n它常在贝叶斯网络中传播消息，估计各节点概率。\n\nProbabilistic-graphical-model\n它是概率图模型里的经典推断算法之一。\n\nVariational Inference\n两者都用于难精确求解时的近似概率推断。\n\nGraph Neural Network\n图神经网络的消息传递思路与它有相似影子。", "relations": { "bayesian-network": { "label": "用于…做推断", "note": "常在贝叶斯网络里传播局部概率。" }, "probabilistic-graphical-model": { "label": "是…经典算法", "note": "它是概率图模型里的核心推断方法。" }, "variational-inference": { "label": "和…同属近似推断", "note": "两者都在难算时近似求后验。" }, "graph-neural-network": { "label": "启发…消息传递", "note": "图神经网络借鉴了节点传消息思路。" } } } } }, { "id": "belief-state", "name": "Belief State", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "markov-decision-process" }, { "to": "world-model" }, { "to": "stateful-agent" }, { "to": "agent-memory" } ], "track": "history", "i18n": { "en": { "fullName": "Belief State", "factExplain": "A probability-based guess about hidden facts in the current situation.", "humanExplain": "Belief state is like playing Clue with half the cards hidden. Your suspect list changes after each clue.\n\nRobots and game AIs use it to decide with missing facts. New clues change the guess.", "humanExplainDisplay": "Belief state is like playing ==Clue==\nwith half the cards hidden.\nYour ==suspect list== changes\nafter each clue.\n\nRobots and game AIs use it\nto decide with missing facts.\nNew clues change the guess.", "relationsNarrative": "MDP\nBelief state often replaces the direct state if the world is partly hidden.\n\nWorld model\nBelief state can be the World model’s current guess about the scene.\n\nStateful agent\nA Stateful agent keeps and updates this guess as things change.\n\nMemory\nMemory stores what happened before. Belief state turns it into the current guess.", "relations": { "markov-decision-process": { "label": "extends … state", "note": "It replaces the plain state if the real state is partly hidden." }, "world-model": { "label": "feeds judgment to …", "note": "It is one way to estimate the current world inside the AI." }, "stateful-agent": { "label": "helps … track the scene", "note": "It lets the agent keep updating its view of the world." }, "agent-memory": { "label": "works with …", "note": "Memory stores past events. Belief state sums up the current guess." } } }, "zh": { "fullName": "信念状态", "factExplain": "对当前环境隐藏信息的概率化内部表示。", "humanExplain": "像相亲局里边聊边打分：谁靠谱、谁端着、谁在画饼，心里那杆秤会一路偷偷改。\n\n常用于信息不全的决策，让系统边观察边修正判断。", "humanExplainDisplay": "像相亲局里边聊边打分：\n谁靠谱、谁端着、谁在画饼，\n心里那杆==秤==\n会一路偷偷==改==。\n\n常用于信息不全的决策，\n让系统边观察边修正判断。", "relationsNarrative": "Markov-decision-process\n当环境状态不可完全观测时，常用它替代直接状态。\n\nWorld model\n它可看作世界模型对当前局势的内部估计。\n\nStateful agent\n有状态代理会持续维护并更新这种局势判断。\n\nMemory\n记忆负责存经历，它负责汇总成当前判断。", "relations": { "markov-decision-process": { "label": "扩展…的状态", "note": "当真实状态看不全时常改用它。" }, "world-model": { "label": "给…提供判断", "note": "它是内部世界估计的一种形式。" }, "stateful-agent": { "label": "支撑…记局势", "note": "让代理持续更新对环境的判断。" }, "agent-memory": { "label": "与…配合使用", "note": "记忆存历史，它管当前判断。" } } } } }, { "id": "bellman-equation", "name": "Bellman Equation", "layer": "L2", "era": "1957", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "markov-decision-process" }, { "to": "q-learning" }, { "to": "dynamic-programming" } ], "track": "history", "i18n": { "en": { "fullName": "Bellman Equation", "factExplain": "A formula that links today’s value to rewards now and later.", "humanExplain": "Bellman Equation is like planning a road trip with snacks. A stop is great if it has donuts now and cookies later.\n\nIt helps AI score a move by adding now and later. You meet it in RL, Q-Learning, and DP.", "humanExplainDisplay": "Bellman Equation is like planning a ==road trip with snacks==.\nA stop is great if it has ==donuts now==\nand cookies later.\n\nIt helps AI score a move\nby adding now and later.\nYou meet it in RL,\nQ-Learning,\nand DP.", "relationsNarrative": "RL\nThe Bellman Equation is the core formula for long-term reward in RL.\n\nMDP\nThe Bellman Equation is usually built on MDP state changes.\n\nQ-Learning\nQ-Learning updates can be seen as sample guesses at the Bellman Equation.\n\nDP\nDP uses the Bellman Equation to compute values or better policies.", "relations": { "reinforcement-learning": { "label": "forms the base of …", "note": "It is the core value formula in RL." }, "markov-decision-process": { "label": "is written on …", "note": "It usually uses MDP states and moves." }, "q-learning": { "label": "supports … updates", "note": "Q-Learning updates are sample guesses at this equation." }, "dynamic-programming": { "label": "helps solve with …", "note": "DP uses it to push values backward step by step." } } }, "zh": { "fullName": "贝尔曼方程", "factExplain": "描述当前价值与未来回报关系的递推方程。", "humanExplain": "像老中医开方子：不能只看今天这口气色，得把后面几服药的走向一并算进去。\n\n常用于计算长期价值，是强化学习评估和找最优策略的基础。", "humanExplainDisplay": "像老中医开方子：\n不能只看今天这口==气色==，\n得把后面几服药的\n==走向==一并算进去。\n\n常用于计算长期价值，\n是强化学习评估和找\n最优策略的基础。", "relationsNarrative": "Reinforcement-learning\n它是强化学习里定义和计算长期回报的核心公式。\n\nMarkov-decision-process\n贝尔曼方程通常建立在 MDP 的状态转移设定上。\n\nQ-Learning\nQ-Learning 的更新规则可看作对它的采样近似。\n\nDynamic-programming\n动态规划常利用它递推求出价值函数或最优策略。", "relations": { "reinforcement-learning": { "label": "构成…基础", "note": "它是强化学习的核心价值公式。" }, "markov-decision-process": { "label": "定义在…上", "note": "通常在 MDP 设定下写出价值关系。" }, "q-learning": { "label": "支撑…更新", "note": "Q 值更新可看作它的近似求解。" }, "dynamic-programming": { "label": "用于…求解", "note": "动态规划常靠它递推最优值。" } } } } }, { "id": "benchmark-contamination", "name": "Benchmark contamination", "layer": "L4", "era": "2023", "publishedAt": "2026-05-28T15:58:23.419Z", "relations": [ { "to": "pretraining" }, { "to": "llm" }, { "to": "agi" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "Benchmark contamination", "factExplain": "A model saw the test questions during training, so its score looks too high.", "humanExplain": "It is like finding tomorrow's spelling test in your backpack. Suddenly you look like a genius, not a wizard.\n\nIt can puff up scores above real skill. You see it on public leaderboards, model tests, and model face-offs.", "humanExplainDisplay": "It is like finding ==tomorrow's spelling test==\nin your backpack.\nSuddenly you ==look like a genius==,\nnot a wizard.\n\nIt can puff up scores\nabove real skill.\nYou see it on public leaderboards,\nmodel tests,\nand model face-offs.", "relationsNarrative": "Pretraining\nIf test data slips into pretraining, later scores stop being trustworthy.\n\nLLM\nBenchmark contamination can make an LLM look smarter than it is.\n\nAGI\nBenchmark contamination can make AGI progress look too rosy.\n\nAI-regulation\nPolluted tests can give AI-regulation a shaky base.", "relations": { "pretraining": { "label": "may include … data", "note": "If test questions enter pretraining, later scores are not trustworthy." }, "llm": { "label": "can misjudge … skill", "note": "Contamination can make an LLM score higher than it should." }, "agi": { "label": "can distort … claims", "note": "Bad benchmark scores can make AGI progress look too strong." }, "ai-regulation": { "label": "can skew … rules", "note": "Rules can go wrong if they use polluted test results." } } }, "zh": { "fullName": "基准污染（Benchmark Contamination）", "factExplain": "模型在训练时见过测试题，会让评测结果失真。", "humanExplain": "基准污染就像考前偷看题库：上榜时像学霸，换套卷子立刻原形毕露。\n\n它会让模型评测虚高，常见于排行榜、第三方评测，也逼着训练数据更仔细清洗。", "humanExplainDisplay": "基准污染就像==考前偷看题库==：\n上榜时像学霸，\n换套卷子立刻==原形毕露==。\n\n它会让模型评测虚高，\n常见于排行榜、第三方评测，\n也逼着训练数据更仔细清洗。", "relationsNarrative": "Pretraining\n如果测试集内容混进预训练数据，后面的评测成绩就不再可靠。\n\nLLM\nbenchmark contamination 会让人高估 LLM 的真实能力，以为它会做，其实只是见过。\n\nAGI\n当大家用基准分数判断 AGI 进展时，污染会让结论显得过度乐观。\n\nAI-regulation\n监管若参考被污染的评测结果，可能建立在失真的能力判断上。", "relations": { "pretraining": { "label": "可能混入…数据", "note": "测试题进入预训练会污染后续评测。" }, "llm": { "label": "会误判…能力", "note": "被污染后，LLM 分数可能虚高。" }, "agi": { "label": "会干扰…判断", "note": "基准失真会夸大通用智能进展。" }, "ai-regulation": { "label": "影响…制定依据", "note": "监管若看错评测，标准也会跑偏。" } } } } }, { "id": "bert", "name": "BERT", "layer": "L3", "era": "2018", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "masked-language-modeling" }, { "to": "embedding" }, { "to": "gpt" } ], "track": "history", "i18n": { "en": { "fullName": "Bidirectional Encoder Representations from Transformers", "factExplain": "A two-way pre-trained language model built on the Transformer encoder.", "humanExplain": "BERT is the careful kid with a worksheet. It checks both sides of the blank before answering.\n\nIt is good at understanding text. You meet it in search, text sorting, and fact pulling.", "humanExplainDisplay": "BERT is the ==careful kid== with a worksheet.\nIt checks ==both sides of the blank== before answering.\n\nIt is good at understanding text.\nYou meet it in search,\ntext sorting,\nand fact pulling.", "relationsNarrative": "Transformer\nBERT is built on the Transformer encoder.\n\nMLM\nBERT trains by hiding some words and guessing them.\n\nEmbedding\nBERT can make meaning vectors for words or sentences.\n\nGPT\nBERT focuses more on text understanding than generation.", "relations": { "transformer": { "label": "is built on …", "note": "BERT is built on the Transformer encoder." }, "masked-language-modeling": { "label": "pre-trains with …", "note": "BERT learns by hiding words and guessing them." }, "embedding": { "label": "creates …", "note": "BERT turns text into meaning vectors for comparison." }, "gpt": { "label": "takes a different path from …", "note": "BERT focuses on understanding, not writing the next words." } } }, "zh": { "fullName": "Bidirectional Encoder Representations from Transformers（双向编码器表征）", "factExplain": "一种基于 Transformer 编码器的双向预训练语言模型。", "humanExplain": "轮到它做题，不会急着往下猜，而是先把前后文都翻一遍，活像阅读理解考场里的稳健选手。\n\n擅长吃透文本意思，常用于分类、检索和信息抽取。", "humanExplainDisplay": "轮到它做题，\n不会急着往下猜，\n而是先把==前后文都翻一遍==，\n活像阅读理解考场里的\n==稳健选手==。\n\n擅长吃透文本意思，\n常用于分类、\n检索和信息抽取。", "relationsNarrative": "Transformer\n它建立在 Transformer 编码器结构之上。\n\nMasked-language-modeling\n它通过遮住部分词再预测来完成预训练。\n\nEmbedding\n它能为句子或词生成可用于比较的语义表示。\n\nGpt\n它更偏文本理解，和主打生成的路线不同。", "relations": { "transformer": { "label": "基于…架构", "note": "它建立在 Transformer 编码器之上。" }, "masked-language-modeling": { "label": "靠…预训练", "note": "它靠遮住词再猜词来学语言。" }, "embedding": { "label": "产出…表示", "note": "它把文本变成可比较的语义向量。" }, "gpt": { "label": "与…路线不同", "note": "它偏理解任务，不主打续写生成。" } } } } }, { "id": "bias-variance-tradeoff", "name": "Bias-Variance Tradeoff", "layer": "L2", "era": "1970s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "empirical-risk-minimization" }, { "to": "statistical-learning-theory" }, { "to": "inductive-bias" } ], "track": "history", "i18n": { "en": { "fullName": "Bias-Variance Tradeoff", "factExplain": "The tradeoff between a model being too simple and too fitted to examples.", "humanExplain": "Memorize every old math answer, and one new question ruins your day. Learn only “math is numbers,” and good luck.\n\nThis tradeoff helps set how bendy a model should be. You meet it when choosing models, tuning settings, and fighting overfitting.", "humanExplainDisplay": "Memorize ==every old math answer==,\nand one new question ruins your day.\nLearn only ==“math is numbers,”==\nand good luck.\n\nThis tradeoff helps set\nhow bendy a model should be.\nYou meet it when choosing models,\ntuning settings,\nand fighting overfitting.", "relationsNarrative": "Regularization\nRegularization helps stop overfitting and balance variance.\n\nERM\nERM can push variance up when it only chases low training error.\n\nSLT\nThis is one of the classic tradeoffs in SLT.\n\nInductive Bias\nA model’s prior assumptions directly change its bias and variance.", "relations": { "regularization": { "label": "balances with …", "note": "Regularization often lowers variance by holding the model back." }, "empirical-risk-minimization": { "label": "covers the blind spot of …", "note": "Only lowering training error can raise variance." }, "statistical-learning-theory": { "label": "is core to …", "note": "It is a classic question in statistical learning." }, "inductive-bias": { "label": "is shaped by …", "note": "Prior assumptions change bias and variance." } } }, "zh": { "fullName": "Bias-Variance Tradeoff（偏差-方差权衡）", "factExplain": "模型在欠拟合与过拟合之间的经典权衡。", "humanExplain": "下棋太死守，容易处处挨打；太爱赌妙手，又可能一招走飞。稳和狠，不能两头都占。\n\n它影响模型该多灵活、管多严，常出现在选模型、调参和防过拟合。", "humanExplainDisplay": "下棋太==死守==，\n容易处处挨打；\n太爱赌妙手，\n又可能一招==走飞==。\n\n稳和狠，不能两头都占。\n它影响模型该多灵活、管多严，\n常出现在选模型、调参\n和防过拟合。", "relationsNarrative": "Regularization\n正则化常用来抑制过拟合，帮助平衡方差。\n\nEmpirical-risk-minimization\n只顾把训练误差压低，常会把方差推高。\n\nStatistical-learning-theory\n它是统计学习理论里最经典的权衡之一。\n\nInductive-bias\n模型的先验假设会直接影响偏差与方差。", "relations": { "regularization": { "label": "靠…做平衡", "note": "正则化常用来压低方差。" }, "empirical-risk-minimization": { "label": "补足…的盲区", "note": "只顾训练集表现，容易失衡。" }, "statistical-learning-theory": { "label": "属于…核心问题", "note": "它是统计学习里的经典命题。" }, "inductive-bias": { "label": "受…影响", "note": "先验假设会改变偏差与方差。" } } } } }, { "id": "big-data", "name": "Big Data", "layer": "L5", "sublayer": "architecture", "era": "1997", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "deep-learning" }, { "to": "pretraining" }, { "to": "distributed-computing" }, { "to": "data-privacy" } ], "track": "history", "i18n": { "en": { "fullName": "Big Data", "factExplain": "Huge data sets that need special tools to store and analyze.", "humanExplain": "Big Data is every receipt from a giant grocery store. One receipt says little. A million show when everyone panic-buys ice cream.\n\nIt helps train AI, suggest videos, and spot fraud. It also makes storage and privacy harder.", "humanExplainDisplay": "Big Data is ==every receipt==\nfrom a giant grocery store.\nOne receipt says little.\n==A million== show\nwhen everyone panic-buys ice cream.\n\nIt helps train AI,\nsuggest videos,\nand spot fraud.\nIt also makes storage\nand privacy harder.", "relationsNarrative": "Deep Learning\nBig Data gives Deep Learning enough examples to learn patterns.\n\nPretraining\nPretraining often starts with massive text and image data.\n\nDistributed Computing\nDistributed Computing splits huge data across many machines.\n\nData-privacy\nMore data means a higher risk of private information leaking.", "relations": { "deep-learning": { "label": "feeds …", "note": "Huge data gives deep models enough examples to learn patterns." }, "pretraining": { "label": "supports …", "note": "Pretraining often starts with massive text and image data." }, "distributed-computing": { "label": "depends on …", "note": "Very large data is often split across many machines." }, "data-privacy": { "label": "raises … risks", "note": "More data means more sensitive information to protect." } } }, "zh": { "fullName": "大数据", "factExplain": "规模巨大、需专门技术处理的数据集合。", "humanExplain": "大数据像气象站的长年记录：单天天气说明不了啥，攒几十年才看出气候往哪走。\n\n它支撑训练、推荐和风控，也放大存储与隐私压力。", "humanExplainDisplay": "大数据像气象站长年记录：\n==单天没戏==，\n攒几十年才看出\n==气候往哪走==。\n\n它支撑训练、推荐和风控，\n也放大存储\n与隐私压力。", "relationsNarrative": "Deep Learning\n大数据为深度学习提供足够训练样本。\n\nPretraining\n预训练常靠海量文本、图像打底。\n\nDistributed Computing\n分布式计算让海量数据能被处理。\n\nData Privacy\n数据越多，隐私泄露风险越高。", "relations": { "deep-learning": { "label": "喂饱…", "note": "海量样本让深度模型学到规律。" }, "pretraining": { "label": "支撑…", "note": "预训练常从海量数据中打底。" }, "distributed-computing": { "label": "依赖…处理", "note": "数据太大，常要分布式拆开算。" }, "data-privacy": { "label": "放大…风险", "note": "数据越多，越要管住敏感信息。" } } } } }, { "id": "bleu", "name": "BLEU", "layer": "L4", "era": "2002", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "machine-translation" }, { "to": "n-gram-language-model" }, { "to": "word-error-rate" }, { "to": "llm-as-a-judge" } ], "track": "history", "i18n": { "en": { "fullName": "Bilingual Evaluation Understudy", "factExplain": "A translation score based on matching short word chunks.", "humanExplain": "BLEU is like a quiz grader with a tiny phrase checklist. Match the word chunks to get points. Skip half the answer and lose points.\n\nIt often scores machine translation and summaries. It is fast, but it does not truly understand meaning.", "humanExplainDisplay": "BLEU is like a quiz grader\nwith a ==tiny phrase checklist==.\nMatch the ==word chunks== to get points.\nSkip half the answer and lose points.\n\nIt often scores machine translation and summaries.\nIt is fast,\nbut it does not truly understand meaning.", "relationsNarrative": "MT\nBLEU was first used to score machine translation quality.\n\nN-gram LM\nBLEU scores text by matching short runs of words.\n\nWER\nBLEU and WER are both automatic scores based on surface matches.\n\nLLM-as-a-judge\nLLM-as-a-judge can help with meaning BLEU misses.", "relations": { "machine-translation": { "label": "scores … quality", "note": "BLEU was first used to score machine translation." }, "n-gram-language-model": { "label": "matches by … chunks", "note": "BLEU mainly compares matching runs of words." }, "word-error-rate": { "label": "also auto-scores like …", "note": "BLEU and WER both judge output by surface matches." }, "llm-as-a-judge": { "label": "is helped by …", "note": "LLM-as-a-judge can check meaning when BLEU cannot." } } }, "zh": { "fullName": "Bilingual Evaluation Understudy，双语评估替补", "factExplain": "用 n-gram 重合度评估译文质量的指标。", "humanExplain": "BLEU 像 KTV 机器打分：歌词对上就加分，少唱一截还要扣分。\n\n常评机器翻译和摘要，出分快但不懂语义好坏。", "humanExplainDisplay": "BLEU 像\n==KTV 机器打分==：\n歌词对上就加分，\n==少唱一截==还扣分。\n\n常评机器翻译和摘要，\n出分快但不懂语义好坏。", "relationsNarrative": "Machine Translation\nBLEU 最早用于自动评估机器翻译输出质量。\n\nN-gram LM\nBLEU 依赖连续词片段的重合度来计分。\n\nWER\nBLEU 和 WER 都是基于表面匹配的自动指标。\n\nLLM-as-a-judge\nLLM-as-a-judge 可补足 BLEU 不懂语义的问题。", "relations": { "machine-translation": { "label": "评估…质量", "note": "BLEU 最早就是为机器翻译打分。" }, "n-gram-language-model": { "label": "按…算重合", "note": "它主要比较连续词片段的重合度。" }, "word-error-rate": { "label": "同属自动评分", "note": "两者都用表面匹配衡量输出差异。" }, "llm-as-a-judge": { "label": "被…补足", "note": "主观质量常需要模型或人工再判断。" } } } } }, { "id": "bm25", "name": "BM25", "layer": "L2", "era": "1994", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "information-retrieval" }, { "to": "rag" }, { "to": "vector-search" } ], "track": "history", "i18n": { "en": { "fullName": "Best Matching 25", "factExplain": "A search method for ranking documents by keyword match.", "humanExplain": "BM25 is like hunting for fries in a school cafeteria. The short “FRIES” sign beats a giant salad poster.\n\nSearch and knowledge bases use it. It finds likely text first. Then the model answers from that text.", "humanExplainDisplay": "BM25 is like hunting for ==fries==\nin a school cafeteria.\nThe short ==“FRIES” sign== beats\na giant salad poster.\n\nSearch and knowledge bases use it.\nIt finds likely text first.\nThen the model answers from that text.", "relationsNarrative": "IR\nBM25 is a classic ranking method in IR.\n\nRAG\nRAG often uses BM25 to find text before the model answers.\n\nVector search\nBM25 pairs with Vector search for word match and meaning match.", "relations": { "information-retrieval": { "label": "is a classic … method", "note": "BM25 is a classic ranking method in IR." }, "rag": { "label": "often retrieves for …", "note": "RAG often uses BM25 to find useful text first." }, "vector-search": { "label": "pairs with …", "note": "Keyword search and meaning search often work better together." } } }, "zh": { "fullName": "Best Matching 25，一种经典关键词检索排序算法", "factExplain": "一种按关键词匹配相关性给文档排序的方法。", "humanExplain": "BM25像在夜市找煎饼摊：招牌对不对、吆喝有没有提到料、废话少不少，越对味越先被你看见。\n\n常用在搜索和知识库检索，先捞相关资料，再交给模型作答。", "humanExplainDisplay": "BM25像在夜市找煎饼摊：\n==招牌对不对==、\n吆喝有没有提到料、\n废话少不少，越对味越==先被看见==。\n\n常用在搜索和知识库检索，\n先捞相关资料，\n再交给模型作答。", "relationsNarrative": "Information Retrieval\nBM25 是信息检索领域的经典排序方法。\n\nRAG\nRAG 常先用它找资料，再让模型基于资料作答。\n\nVector Search\n它擅长关键词匹配，常与语义检索配合使用。", "relations": { "information-retrieval": { "label": "属于…经典方法", "note": "它是传统信息检索里的代表排序算法。" }, "rag": { "label": "常为…做召回", "note": "RAG 常先靠它找出相关文本。" }, "vector-search": { "label": "与…形成互补", "note": "关键词匹配和语义检索常一起用。" } } } } }, { "id": "boltzmann-machine", "name": "Boltzmann Machine", "layer": "L3", "era": "1985", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "hopfield-network" }, { "to": "markov-chain-monte-carlo" }, { "to": "latent-variable-model" } ], "track": "history", "i18n": { "en": { "fullName": "Boltzmann Machine", "factExplain": "A random neural network using energy scores to model data patterns.", "humanExplain": "A Boltzmann Machine is like a school cafeteria seating shuffle. Everyone keeps swapping seats until the table feels less awkward.\n\nIt learns hidden patterns from unlabeled data. Training often uses sampling, so it tries many possible states.", "humanExplainDisplay": "A Boltzmann Machine is like\n==a school cafeteria seating shuffle==.\nEveryone keeps swapping seats\nuntil the table feels ==less awkward==.\n\nIt learns hidden patterns\nfrom unlabeled data.\nTraining often uses sampling,\nso it tries many possible states.", "relationsNarrative": "Energy-Based Model\nA Boltzmann Machine is a classic Energy-Based Model.\n\nHopfield Network\nA Boltzmann Machine grew from the Hopfield Network and adds randomness.\n\nMCMC\nA Boltzmann Machine often uses MCMC to sample during training.\n\nLatent Model\nA Boltzmann Machine can be a Latent Model with hidden units.", "relations": { "hopfield-network": { "label": "grew from …", "note": "It adds randomness, so it can create samples." }, "markov-chain-monte-carlo": { "label": "trains with …", "note": "MCMC helps estimate the patterns during training." }, "latent-variable-model": { "label": "can act as a …", "note": "Hidden units stand for factors we cannot see." } } }, "zh": { "fullName": "玻尔兹曼机", "factExplain": "一种用能量函数建模数据分布的随机神经网络。", "humanExplain": "玻尔兹曼机像麻将桌听牌：牌友互相牵动，试来试去，最后摸到低能好局。\n\n用于无监督建模和特征学习，训练常靠采样。", "humanExplainDisplay": "玻尔兹曼机像麻将桌听牌：\n牌友==互相牵动==，\n试来试去，\n最后摸到==低能好局==。\n\n用于无监督建模\n和特征学习，\n训练常靠采样。", "relationsNarrative": "Energy-based Model\n玻尔兹曼机是早期能量模型代表，用能量定义概率。\n\nHopfield Network\n它像加入随机性的 Hopfield Network，可生成样本。\n\nMCMC\n训练常用 MCMC 采样估计难算的分布。\n\nLatent Model\n隐藏单元让它能表示数据背后的潜在因素。", "relations": { "hopfield-network": { "label": "从…发展而来", "note": "加入随机性，能生成样本。" }, "markov-chain-monte-carlo": { "label": "用…采样训练", "note": "训练常靠采样估计分布。" }, "latent-variable-model": { "label": "可作为…", "note": "隐藏单元表示未观测因素。" } } } } }, { "id": "brain-computer-interface", "name": "BCI", "layer": "L6", "era": "1973", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "ai-device-ai" }, { "to": "affective-computing" }, { "to": "human-in-the-loop" }, { "to": "embodied-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Brain-Computer Interface", "factExplain": "A technology that sends brain signals to a computer or machine.", "humanExplain": "BCI is like giving your brain a tiny TV remote. Your hand stays still, but the gadget gets the “go” signal.\n\nYou meet it in prosthetics and rehab. It can help people communicate, but it is not mind reading.", "humanExplainDisplay": "BCI is like giving your brain\n==a tiny TV remote==.\nYour hand stays still,\nbut the gadget gets the ==“go” signal==.\n\nYou meet it in prosthetics and rehab.\nIt can help people communicate,\nbut it is not mind reading.", "relationsNarrative": "AI device\nBCI could become a new device input, from fingers to brain signals.\n\nAffective Computing\nBrain signals can help spot emotion, attention, and tiredness.\n\nHuman-in-the-loop\nBCI lets human intent enter the system loop more directly.\n\nEmbodied AI\nBCI can turn neural intent into prosthetic or robot movement.", "relations": { "ai-device-ai": { "label": "becomes a new input for …", "note": "Brain signals could become a new way to control devices." }, "affective-computing": { "label": "feeds signals to …", "note": "Brain waves can help spot emotion and attention." }, "human-in-the-loop": { "label": "makes … more direct", "note": "Human intent can enter the system loop more directly." }, "embodied-ai": { "label": "drives action in …", "note": "Neural intent can turn into machine movement." } } }, "zh": { "fullName": "脑机接口", "factExplain": "连接大脑信号与计算机的交互技术。", "humanExplain": "脑机接口像给大脑插上游戏手柄：手没动，机器已收到开跑指令。\n\n用于假肢、康复和辅助沟通，但还不是读心术。", "humanExplainDisplay": "脑机接口像给大脑\n==插上游戏手柄==：\n手没动，\n机器已收到==开跑指令==。\n\n用于假肢、康复\n和辅助沟通，\n但还不是读心术。", "relationsNarrative": "AI Device\nBCI 可能成为设备新入口，从手指输入到脑信号。\n\nAffective Computing\n脑电信号可辅助识别情绪、注意力和疲劳。\n\nHuman-in-the-loop\n它让人的意图更直接进入系统回路。\n\nEmbodied AI\n它可把神经意图转成假肢或机器人动作。", "relations": { "ai-device-ai": { "label": "作为…新入口", "note": "脑信号可能成为下一代设备输入。" }, "affective-computing": { "label": "给…提供信号", "note": "脑电可辅助识别情绪与注意状态。" }, "human-in-the-loop": { "label": "让…更直接", "note": "人类意图可直接进入系统回路。" }, "embodied-ai": { "label": "驱动…行动", "note": "可把神经意图转成机器动作。" } } } } }, { "id": "byte-pair-encoding", "name": "BPE", "layer": "L2", "era": "2016", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "embedding" }, { "to": "llm" } ], "track": "history", "i18n": { "en": { "fullName": "Byte Pair Encoding", "factExplain": "A way to split text into smaller word pieces called tokens.", "humanExplain": "BPE is like making words with fridge magnets. Common chunks stay stuck, and weird new words still get built.\n\nIt keeps the AI’s word list small. Many LLMs use it before reading your prompt.", "humanExplainDisplay": "BPE is like making words with ==fridge magnets==.\n==Common chunks== stay stuck,\nand weird new words still get built.\n\nIt keeps the AI’s word list small.\nMany LLMs use it before reading your prompt.", "relationsNarrative": "Token\nBPE decides which tokens a piece of text becomes.\n\nEmbedding\nBPE splits text first, then each piece can get an embedding.\n\nLLM\nMany LLMs use BPE as their basic text splitter.", "relations": { "token": { "label": "creates …", "note": "It decides which tokens a piece of text becomes." }, "embedding": { "label": "comes before …", "note": "Text is split first, then each piece can get an embedding." }, "llm": { "label": "splits text for …", "note": "Many LLMs use BPE to handle input text." } } }, "zh": { "fullName": "字节对编码", "factExplain": "一种把文本切成子词 token 的编码方法。", "humanExplain": "BPE 像乐高拼句子：常同框的小块先扣牢，生词也能临时拼出来。\n\n它控制词表并减少生词，是大模型分词常用底座。", "humanExplainDisplay": "BPE 像\n==乐高拼句子==：\n常同框的小块先扣牢，\n生词也能临时拼出来。\n\n它控制词表并减少生词，\n是大模型分词，\n常用底座。", "relationsNarrative": "Token\nBPE 决定一段文本会被切成哪些 token。\n\nEmbedding\n文本先经 BPE 切块，才能进入向量表示。\n\nLLM\n许多大语言模型用 BPE 作为分词底座。", "relations": { "token": { "label": "生成…", "note": "它决定文本会被切成哪些 token。" }, "embedding": { "label": "先于…", "note": "文本先被切块，才能查向量表示。" }, "llm": { "label": "服务…分词", "note": "许多大模型用它处理输入文本。" } } } } }, { "id": "canny-edge-detector", "name": "Canny Edge Detector", "layer": "L2", "era": "1986", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "hough-transform" }, { "to": "image-segmentation" } ], "track": "history", "i18n": { "en": { "fullName": "Canny Edge Detector", "factExplain": "An algorithm that extracts image edges through several careful steps.", "humanExplain": "Canny is like a very neat kid with a coloring book. It ignores the smudges, then traces the real outlines in black.\n\nIt is a classic Computer Vision tool. It helps programs see borders before segmentation or line finding.", "humanExplainDisplay": "Canny is like a ==very neat kid==\nwith a coloring book.\nIt ignores the smudges,\nthen traces the ==real outlines== in black.\n\nIt is a classic Computer Vision tool.\nIt helps programs see borders\nbefore segmentation or line finding.", "relationsNarrative": "Computer Vision\nCanny is a classic edge detector in Computer Vision.\n\nHough Transform\nCanny often finds edges first, then Hough Transform finds lines or circles.\n\nImage Segmentation\nCanny edge maps can help Image Segmentation find object borders.", "relations": { "computer-vision": { "label": "is a classic in …", "note": "It is a classic edge tool for image understanding." }, "hough-transform": { "label": "sets up …", "note": "Edges come first, so Hough can vote for lines or circles." }, "image-segmentation": { "label": "helps … find borders", "note": "Edge lines can show where one region ends." } } }, "zh": { "fullName": "Canny 边缘检测器", "factExplain": "一种用多阶段处理提取图像边缘的算法。", "humanExplain": "Canny 像装修师傅弹墨线：先扫掉浮灰，再把墙角轮廓一笔勒清。\n\n用于传统视觉，帮分割和检测先看清边界。", "humanExplainDisplay": "Canny 像装修师傅==弹墨线==：\n先扫掉浮灰，\n再把墙角轮廓\n==一笔勒清==。\n\n用于传统视觉，\n帮分割和检测\n先看清边界。", "relationsNarrative": "Computer Vision\n它是传统计算机视觉里的经典边缘检测算法。\n\nHough Transform\n它常先找边缘，再让 Hough Transform 找直线或圆。\n\nImage Segmentation\n边缘图可辅助分割判断物体边界。", "relations": { "computer-vision": { "label": "属于…经典算法", "note": "经典边缘算法，服务图像理解。" }, "hough-transform": { "label": "为…找线打底", "note": "先有边缘，才好投票找线。" }, "image-segmentation": { "label": "辅助…定边界", "note": "边缘线能提示区域分界。" } } } } }, { "id": "captcha", "name": "CAPTCHA", "layer": "L6", "era": "2003", "publishedAt": "2026-05-30T03:10:23.228Z", "relations": [ { "to": "turing-test" }, { "to": "agent" }, { "to": "computer-use" } ], "track": "today", "i18n": { "en": { "fullName": "Completely Automated Public Turing test to tell Computers and Humans Apart", "factExplain": "A check used to tell real people from bots.", "humanExplain": "CAPTCHA is the website bouncer with a pop quiz. It asks for traffic lights or twisty letters before you enter.\n\nYou meet it when you log in or sign up, and ticket sites use it too. It blocks bots, but it can trip up real people.", "humanExplainDisplay": "CAPTCHA is the ==website bouncer== with a ==pop quiz==.\nIt asks for traffic lights\nor twisty letters\nbefore you enter.\n\nYou meet it when you log in or sign up,\nand ticket sites use it too.\nIt blocks bots,\nbut it can trip up real people.", "relationsNarrative": "Turing-test\nCAPTCHA flips the Turing-test and makes you prove you are human.\n\nAgent\nMany sites use CAPTCHA to stop an Agent from signing up or buying tickets at scale.\n\nComputer use\nWhen AI uses a web page for you, CAPTCHA is often the first roadblock.", "relations": { "turing-test": { "label": "flips the …", "note": "A Turing-test asks if a machine seems human, but CAPTCHA checks you." }, "agent": { "label": "blocks …", "note": "Many sites use CAPTCHA to stop Agents from acting at scale." }, "computer-use": { "label": "sets a hurdle for …", "note": "CAPTCHA is often the first hurdle for AI using a website." } } }, "zh": { "fullName": "验证码", "factExplain": "一种用来区分真人和机器的身份验证方式。", "humanExplain": "验证码像小区门口保安：让你认红绿灯，不是为难你，是怕机器人混进业主群。\n\n它用于登录、注册和抢票防刷；AI 变强后，题也越来越刁。", "humanExplainDisplay": "验证码像==小区门口保安==：\n让你认红绿灯，\n不是为难你，\n是怕==机器人混进业主群==。\n\n它用于登录、注册和抢票防刷；\nAI 变强后，\n题也越来越刁。", "relationsNarrative": "Turing-test\n图灵测试看机器像不像人，验证码反过来——要你证明自己不是机器。\n\nAgent\n很多网站会用 CAPTCHA 限制 Agent 批量注册、刷票或自动提交操作。\n\nComputer use\nAI 代你点网页填表单时，验证码常是第一道拦路坎。", "relations": { "turing-test": { "label": "像…的反向版", "note": "图灵测试测像不像人，它测是不是人。" }, "agent": { "label": "用来拦住…", "note": "很多网站用验证码限制 Agent 自动操作。" }, "computer-use": { "label": "给…设门槛", "note": "验证码常是电脑代操作的第一道坎。" } } } } }, { "id": "cart", "name": "CART", "layer": "L2", "era": "1984", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "regression" }, { "to": "gradient-boosting" } ], "track": "history", "i18n": { "en": { "fullName": "Classification and Regression Tree", "factExplain": "A model that predicts by splitting data with tree-like rules.", "humanExplain": "CART is a strict teacher with a seating chart. Glasses to the left, chatterboxes to the back, and somehow everyone gets a seat.\n\nIn AI, it asks simple questions to sort things or predict numbers, like prices. Its rules are easy to read, but a very deep tree can memorize the homework.", "humanExplainDisplay": "CART is a ==strict teacher== with a seating chart.\n==Glasses to the left==,\nchatterboxes to the back,\nand somehow everyone gets a seat.\n\nIn AI, it asks simple questions\nto sort things or predict numbers,\nlike prices.\nIts rules are easy to read,\nbut a very deep tree can memorize the homework.", "relationsNarrative": "Classification\nCART often puts examples into different groups.\n\nRegression\nCART can also predict numbers, like house prices.\n\nGradient Boosting\nGradient Boosting often stacks many CART trees.", "relations": { "classification": { "label": "can do … tasks", "note": "It can put examples into different groups." }, "regression": { "label": "can also do …", "note": "It can predict a number, like a price or score." }, "gradient-boosting": { "label": "often powers …", "note": "Many boosting methods use CART as the base tree." } } }, "zh": { "fullName": "分类与回归树", "factExplain": "一种用树形分裂规则做预测的模型。", "humanExplain": "CART 像果园的分拣传送带：先按大小掉进不同口，再按甜不甜分箱，一层层定下每个果子该去哪。\n\n常用于分类和数值预测；规则直观，但树太深易过拟合。", "humanExplainDisplay": "CART 像果园的分拣传送带：\n先按==大小==掉进不同口，\n再按甜不甜分箱，\n一层层定下\n每个果子==该去哪==。\n\n常用于分类和数值预测；\n规则直观，\n但树太深易过拟合。", "relationsNarrative": "Classification\n它最常见的用途之一，就是把样本分到不同类别。\n\nRegression\n除了分类型任务，它也能预测房价这类连续数值。\n\nGradient Boosting\n梯度提升常把它作为基础弱学习器反复叠加。", "relations": { "classification": { "label": "可用于…任务", "note": "它能把样本分到不同类别。" }, "regression": { "label": "也可做…预测", "note": "它也能预测连续数值结果。" }, "gradient-boosting": { "label": "常作为…基树", "note": "很多提升方法都拿它当底座。" } } } } }, { "id": "causal-inference", "name": "Causal Inference", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "structural-causal-model" }, { "to": "do-calculus" }, { "to": "regression" }, { "to": "bayesian-network" } ], "track": "history", "i18n": { "en": { "fullName": "Causal Inference", "factExplain": "A method for finding cause-and-effect links in data.", "humanExplain": "Causal inference is detective work after the office Wi‑Fi dies. It asks who unplugged the router, not who yelled first.\n\nDoctors, economists, and policy teams use it. It stops us from calling timing a cause.", "humanExplainDisplay": "Causal inference is ==detective work==\nafter the office Wi‑Fi dies.\nIt asks ==who unplugged the router==,\nnot who yelled first.\n\nDoctors, economists, and policy teams use it.\nIt stops us from calling timing a cause.", "relationsNarrative": "SCM\nSCM is a common formal way to write causal inference.\n\nDo-Calculus\nDo-Calculus helps it predict what changes after an intervention.\n\nRegression\nRegression finds statistical links, but it does not prove cause by itself.\n\nBayesian Network\nCausal inference and Bayesian Network both use diagrams for variable links.", "relations": { "structural-causal-model": { "label": "is often written as …", "note": "SCM writes causal ideas as diagrams and equations." }, "do-calculus": { "label": "studies interventions with …", "note": "Do-Calculus helps predict results after an action is forced." }, "regression": { "label": "is not the same as …", "note": "Regression can find links, but not always causes." }, "bayesian-network": { "label": "looks similar to …", "note": "Both use diagrams, but they answer different questions." } } }, "zh": { "fullName": "因果推断", "factExplain": "从数据中识别原因与结果关系的方法。", "humanExplain": "因果推断像查楼里漏水：光看谁家地湿没用，得顺着水管和时间，找出到底哪截先破的。\n\n常用于医疗、经济、政策分析；避免把同时发生误当因果。", "humanExplainDisplay": "因果推断像查楼里漏水：\n光看==谁家地湿==没用，\n得顺着水管和时间，\n找出到底==哪截先破的==。\n\n常用于医疗、经济、\n政策分析；\n避免把同时发生误当因果。", "relationsNarrative": "Structural Causal Model\n结构因果模型是因果推断最常见的形式化表达。\n\nDo-Calculus\nDo-Calculus 帮它推导干预发生后的结果变化。\n\nRegression\n回归擅长找统计关联，但不能自动说明因果。\n\nBayesian Network\n它和贝叶斯网络都用图表示变量关系。", "relations": { "structural-causal-model": { "label": "常用…表达", "note": "它常借助因果图和结构方程建模。" }, "do-calculus": { "label": "用…做干预分析", "note": "Do-Calculus 用来推导干预后的结果。" }, "regression": { "label": "不等于…", "note": "回归能找关联，不一定能定因果。" }, "bayesian-network": { "label": "与…形式相近", "note": "两者都画图，但目标不完全一样。" } } } } }, { "id": "chain-of-thought", "name": "Chain-of-thought", "layer": "L2", "era": "2022", "publishedAt": "2026-05-23T09:25:00Z", "relations": [ { "to": "prompt" }, { "to": "reasoning-model" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Chain-of-thought", "factExplain": "A way to guide AI to solve hard problems step by step.", "humanExplain": "Chain-of-thought is the AI showing its math homework. No more shouting “42!” from the couch.\n\nIt asks the model to solve a problem in steps. This can help with hard tasks, but tidy steps can still be pretend.", "humanExplainDisplay": "Chain-of-thought is the AI\n==showing its math homework==.\nNo more shouting ==“42!” from the couch==.\n\nIt asks the model to solve a problem in steps.\nThis can help with hard tasks,\nbut tidy steps can still be pretend.", "relationsNarrative": "Prompt\nA Prompt can ask Chain-of-thought to unfold the middle steps.\n\nReasoning-model\nReasoning-models often improve Chain-of-thought style reasoning.\n\nLLM\nChain-of-thought can help an LLM handle hard problems more steadily.", "relations": { "prompt": { "label": "guided by …", "note": "A Prompt can ask the model to show middle steps." }, "reasoning-model": { "label": "core skill for …", "note": "Reasoning-models often strengthen step-by-step thinking." }, "llm": { "label": "helps … think step by step", "note": "Chain-of-thought can make an LLM steadier on hard problems." } } }, "zh": { "fullName": "思维链", "factExplain": "引导模型分步骤推理以解决复杂问题的方法或过程。", "humanExplain": "思维链像让 AI 别光报答案，先把草稿纸摊出来算一算。\n\n它能提升复杂题表现，但草稿不等于真理，写得像推理也可能是演的。", "humanExplainDisplay": "思维链像让 AI 别急着报答案，\n先把==草稿纸==摊出来。\n一步一步算，别靠气势赢。\n\n它能提升复杂题表现。\n但写得像推理，\n不代表一定真想明白了。", "relationsNarrative": "Prompt\nChain-of-thought 通过 Prompt 要求模型展开中间步骤。\n\nReasoning-model\nReasoning-model 往往强化了 Chain-of-thought 式推理。\n\nLLM\nChain-of-thought 能提升 LLM 处理复杂问题的稳定性。", "relations": { "prompt": { "label": "由…引导" }, "reasoning-model": { "label": "是…的核心能力" }, "llm": { "label": "让…分步思考" } } } } }, { "id": "chat-template", "name": "Chat template", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "prompt" }, { "to": "system-prompt" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Chat template", "factExplain": "A rule for arranging chat messages in the format a chat model expects.", "humanExplain": "Same words, very different vibe: lunchbox note or detention slip. A chat template is the little form with all the boxes.\n\nIt puts the system message and user message in the right spots. Different LLMs expect different forms. This can change answers and compatibility.", "humanExplainDisplay": "Same words, very different vibe:\n==lunchbox note== or ==detention slip==.\nA chat template is\nthe little form with all the boxes.\n\nIt puts the system message\nand user message in the right spots.\nDifferent LLMs expect different forms.\nThis can change answers and compatibility.", "relationsNarrative": "Prompt\nA chat template decides how the prompt becomes chat input.\n\nSystem prompt\nA chat template decides which message spot holds the system prompt.\n\nLLM\nDifferent LLMs often use different chat templates.", "relations": { "prompt": { "label": "sets the format for …", "note": "It wraps the prompt in a shape the model understands." }, "system-prompt": { "label": "places …", "note": "The system prompt usually needs a fixed message spot." }, "llm": { "label": "matches … input", "note": "Different LLMs often need different chat formats." } } }, "zh": { "fullName": "聊天模板", "factExplain": "把消息按固定格式组织给聊天模型的输入规则。", "humanExplain": "同一句话，写在请假条、检讨书还是情书上，味道完全不同；它管的就是这个版式。\n\n它规定系统话和用户话怎么装进模型，常影响兼容性与回答稳定性。", "humanExplainDisplay": "同一句话，\n写在==请假条、检讨书还是情书==上，\n味道完全不同；\n它管的就是这个版式。\n\n它规定系统话和用户话\n怎么装进模型，\n常影响兼容性与回答稳定性。", "relationsNarrative": "Prompt\n它规定提示词如何包装成聊天输入格式。\n\nSystem prompt\n它决定系统指令放在哪个消息位置。\n\nLLM\n不同聊天模型往往对应不同聊天模板。", "relations": { "prompt": { "label": "规定…格式", "note": "它把提示词包装成模型认得的样子。" }, "system-prompt": { "label": "安放…位置", "note": "系统指令通常要放在固定消息位。" }, "llm": { "label": "适配…输入", "note": "不同聊天模型常有不同消息格式。" } } } } }, { "id": "chatbot", "name": "Chatbot", "layer": "L4", "era": "1966", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "chatgpt" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Chatbot", "factExplain": "An AI program you talk to through messages.", "humanExplain": "A chatbot is your group-chat friend with no bedtime. Text it at 3 a.m. It replies before your pizza rolls get hot.\n\nYou meet it in customer support, small talk, and quick Q&A. It is often the front desk of an AI product.", "humanExplainDisplay": "A chatbot is your ==group-chat friend==\nwith no bedtime.\nText it at 3 a.m.\nIt replies before your ==pizza rolls get hot==.\n\nYou meet it in customer support,\nsmall talk,\nand quick Q&A.\nIt is often the front desk\nof an AI product.", "relationsNarrative": "ELIZA\nELIZA was a classic early chatbot.\n\nLLM\nModern chatbots often use an LLM to understand and write replies.\n\nChatGPT\nChatGPT brought chatbots to everyday people.\n\nAgent\nAn Agent can plan and do tasks beyond chatting.", "relations": { "llm": { "label": "writes replies with …", "note": "Modern chatbots often use an LLM to answer." }, "chatgpt": { "label": "went mainstream through …", "note": "ChatGPT made chatbots feel normal to many people." }, "agent": { "label": "can grow into …", "note": "An Agent can go beyond chat and take actions." } } }, "zh": { "fullName": "聊天机器人", "factExplain": "通过对话与用户交互的 AI 程序。", "humanExplain": "聊天机器人像楼下便利店夜班店员：半夜三点随叫随答，准不准另算。\n\n常做客服、陪聊和问答，是 AI 产品门面。", "humanExplainDisplay": "聊天机器人像楼下便利店\n==夜班店员==：\n半夜三点==随叫随答==，\n准不准另算。\n\n常做客服、陪聊和问答，\n是 AI 产品门面。", "relationsNarrative": "ELIZA\nELIZA 是早期聊天机器人的经典代表。\n\nLLM\n现代聊天机器人常用 LLM 理解并生成回复。\n\nChatGPT\nChatGPT 让聊天机器人真正走向大众。\n\nAgent\nAgent 可在聊天之外继续规划和执行任务。", "relations": { "llm": { "label": "用…生成回复", "note": "现代聊天机器人常由 LLM 驱动。" }, "chatgpt": { "label": "被…带火", "note": "ChatGPT 让聊天机器人走向大众。" }, "agent": { "label": "可升级成…", "note": "Agent 比聊天机器人更会行动。" } } } } }, { "id": "chatgpt-work", "name": "ChatGPT Work", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "chatgpt" }, { "to": "ai-office-automation" }, { "to": "enterprise-ai-deployment" }, { "to": "kimi-work" } ], "track": "today", "i18n": { "en": { "fullName": "ChatGPT Work", "factExplain": "An agent inside ChatGPT that gathers context across your apps and delivers finished work.", "humanExplain": "ChatGPT Work is ChatGPT with a real job. You name the goal. It digs through your apps, does the steps, and hands in finished work.\n\nYou meet it in office tasks across many apps. It can run for hours on its own. Permissions still need a lock.", "humanExplainDisplay": "ChatGPT Work is ChatGPT\n==with a real job==.\nYou name the goal.\nIt digs through your apps,\ndoes the steps,\nand ==hands in finished work==.\n\nYou meet it in office tasks\nacross many apps.\nIt can run for hours on its own.\nPermissions still need a lock.", "relationsNarrative": "ChatGPT\nChatGPT Work turns ChatGPT into an agent that finishes office work.\n\nAI Office Automation\nChatGPT Work helps automate docs, sheets, and meetings.\n\nEnterprise AI Deployment\nChatGPT Work needs clear access rules and data boundaries.\n\nKimi Work\nChatGPT Work and Kimi Work both want the AI office doorway.", "relations": { "chatgpt": { "label": "brings … to work", "note": "It turns ChatGPT into an agent that does office work." }, "ai-office-automation": { "label": "pushes …", "note": "It handles docs, sheets, and meetings at work." }, "enterprise-ai-deployment": { "label": "needs …", "note": "Company use needs access rules and data boundaries." }, "kimi-work": { "label": "competes with …", "note": "Both aim to be the doorway for AI office work." } } }, "zh": { "fullName": "ChatGPT 工作版", "factExplain": "ChatGPT 内置的职场智能体，能跨应用收集上下文、拆解步骤，独立完成任务并交付成品文档。", "humanExplain": "ChatGPT Work像给聊天机器人转正上岗：不再只陪聊，接了活自己翻应用、拆步骤，交上来的直接是成品。\n\n用于跨应用办公任务，能独立跑几小时，权限还是要管。", "humanExplainDisplay": "ChatGPT Work像给\n聊天机器人==转正上岗==：\n不再只陪聊，\n接了活自己翻应用、拆步骤，\n交上来的==直接是成品==。\n\n用于跨应用办公任务，\n能独立跑几小时，\n权限还是要管。", "relationsNarrative": "ChatGPT\n它把 ChatGPT 从陪聊升级成能独立干活的职场智能体。\n\nAI Office Automation\n它常用于文档、表格、会议等办公任务自动化。\n\nEnterprise AI Deployment\n企业落地时，权限、数据边界和管理能力很关键。\n\nKimi Work\n两者都在争夺 AI 办公工作流入口。", "relations": { "chatgpt": { "label": "把…带进办公", "note": "它把 ChatGPT 升级成能独立干活的职场智能体。" }, "ai-office-automation": { "label": "推动…", "note": "它常用于文档、表格和会议处理。" }, "enterprise-ai-deployment": { "label": "依赖…落地", "note": "企业使用要处理权限和数据边界。" }, "kimi-work": { "label": "对标…", "note": "两者都瞄准 AI 办公工作流。" } } } } }, { "id": "chatgpt", "name": "ChatGPT", "layer": "L5", "sublayer": "product", "era": "2022", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "rlhf" }, { "to": "prompt" }, { "to": "api" } ], "track": "history", "i18n": { "en": { "fullName": "ChatGPT", "factExplain": "A chat AI product built on a large language model.", "humanExplain": "ChatGPT is like a group chat buddy who never sleeps. Ask a messy question. It comes back with a neat answer.\n\nPeople use it for questions and writing. They also use it to sort notes. It is handy. But it can sound sure and still be wrong.", "humanExplainDisplay": "ChatGPT is like a ==group chat buddy==\nwho ==never sleeps==.\nAsk a messy question.\nIt comes back with a neat answer.\n\nPeople use it for questions and writing.\nThey also use it to sort notes.\nIt is handy.\nBut it can sound sure\nand still be wrong.", "relationsNarrative": "LLM\nChatGPT wraps an LLM as a chat product.\n\nRLHF\nRLHF helps ChatGPT answer in a more human way.\n\nPrompt\nYour prompt points ChatGPT toward its answer.\n\nAPI\nAn API lets other apps use ChatGPT too.", "relations": { "llm": { "label": "puts … in a chat box", "note": "It turns an LLM into a product anyone can use." }, "rlhf": { "label": "chats better with …", "note": "RLHF helps its answers feel more like normal talk." }, "prompt": { "label": "is steered by …", "note": "Your prompt strongly shapes what it says next." }, "api": { "label": "connects through …", "note": "Apps can call ChatGPT through an API, not just chat." } } }, "zh": { "fullName": "聊天生成式预训练变换器产品", "factExplain": "基于大语言模型的对话式 AI 产品。", "humanExplain": "像班里那个永远不下线的学霸同桌：你题刚递过去，它先给思路，转头还能替你润作文、列提纲、补资料。\n\n常用于问答、写作和信息整理；顺手好用，但偶尔也会自信答偏。", "humanExplainDisplay": "像班里那个\n永远不下线的==学霸同桌==：\n你题刚递过去，它先给思路，\n转头还能==替你润作文==、列提纲、补资料。\n\n常用于问答、写作和信息整理；\n顺手好用，但偶尔也会自信答偏。", "relationsNarrative": "LLM\n它本质上是把大语言模型包装成对话产品。\n\nRLHF\n人类反馈训练让它更会按人话方式回答。\n\nPrompt\n用户给出的提示词，决定它往哪儿答。\n\nAPI\n除了自己聊天，也能通过接口接入别的产品。", "relations": { "llm": { "label": "把…装进聊天框", "note": "它把大语言模型变成人人可用产品。" }, "rlhf": { "label": "靠…更会聊天", "note": "人类反馈让它回答更像正常对话。" }, "prompt": { "label": "通过…被驱动", "note": "用户怎么提问，会直接影响输出。" }, "api": { "label": "还能经由…接入", "note": "除了聊天界面，也能被应用调用。" } } } } }, { "id": "chinese-ai-models-going-global", "name": "Chinese AI Models Going Global", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "deepseek" }, { "to": "qwen" }, { "to": "open-source-model" }, { "to": "ai-export-controls" } ], "track": "today", "i18n": { "en": { "fullName": "Chinese AI Models Going Global", "factExplain": "The spread of Chinese AI models into overseas markets and developer communities.", "humanExplain": "Chinese AI models going global is like a new taco truck outside the school cafeteria. Suddenly everyone checks the menu, and the old lunch line gets nervous.\n\nIt changes model rankings and launch plans overseas. You see it on cloud platforms and in open-source communities.", "humanExplainDisplay": "Chinese AI models going global is like\na ==new taco truck== outside the school cafeteria.\nSuddenly everyone checks the menu,\nand the ==old lunch line gets nervous==.\n\nIt changes model rankings\nand launch plans overseas.\nYou see it on cloud platforms\nand in open-source communities.", "relationsNarrative": "DeepSeek\nDeepSeek is a breakout example of Chinese AI models going global.\n\nQwen\nQwen reaches overseas developers through open weights and ecosystem tools.\n\nOpen-source-model\nOpen-source models are a key path into the global AI ecosystem.\n\nAI Export Controls\nAI Export Controls can affect compute, market access, and partnerships.", "relations": { "deepseek": { "label": "breaks out through …", "note": "DeepSeek made global users notice Chinese models with low-cost results." }, "qwen": { "label": "spreads its ecosystem through …", "note": "Qwen reaches overseas developers through open weights and tools." }, "open-source-model": { "label": "lowers barriers with …", "note": "Open weights make overseas trials easier." }, "ai-export-controls": { "label": "is limited by …", "note": "Geopolitical rules can slow global rollout." } } }, "zh": { "fullName": "中国 AI 模型出海", "factExplain": "中国 AI 模型进入海外市场和生态的趋势。", "humanExplain": "中国模型出海像夜市新开一摊：东西又好又便宜，排队的人呼啦全过去，老摊得重新琢磨怎么留客。\n\n它影响海外模型评测和部署，常见于云平台、开源社区。", "humanExplainDisplay": "中国模型出海像\n==夜市新开一摊==：\n东西又好又便宜，\n排队的人呼啦全过去，\n老摊得==重新琢磨留客==。\n\n它影响海外模型评测和部署，\n常见于云平台、开源社区。", "relationsNarrative": "DeepSeek\nDeepSeek 是中国模型出海最出圈的代表之一。\n\nQwen\nQwen 通过开源权重和生态工具触达海外开发者。\n\nOpen-source Model\n开源模型是中国模型进入全球生态的重要通道。\n\nAI Export Controls\n出口管制会影响算力、市场准入和合作边界。", "relations": { "deepseek": { "label": "以…破圈", "note": "低成本表现让全球用户注意到中国模型。" }, "qwen": { "label": "以…扩散生态", "note": "开源生态帮助它进入海外社区。" }, "open-source-model": { "label": "借…降低门槛", "note": "开放权重能降低海外试用门槛。" }, "ai-export-controls": { "label": "受…约束", "note": "地缘规则会改变出海速度。" } } } } }, { "id": "chinese-room-argument", "name": "Chinese Room Argument", "layer": "L1", "era": "1980", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "turing-test" }, { "to": "symbolic-ai" }, { "to": "ai-consciousness-debate" }, { "to": "physical-symbol-system-hypothesis" } ], "track": "history", "i18n": { "en": { "fullName": "Chinese Room Argument", "factExplain": "A thought experiment saying symbol rules are not the same as real understanding.", "humanExplain": "It is like answering Mandarin texts with a giant cheat sheet. Your replies look smooth, but your brain is just matching squiggles.\n\nIt challenges chat-style tests for AI. A good reply may show rules, not understanding or a mind.", "humanExplainDisplay": "It is like answering Mandarin texts\nwith a ==giant cheat sheet==.\nYour replies look smooth,\nbut your brain is just ==matching squiggles==.\n\nIt challenges chat-style tests for AI.\nA good reply may show rules,\nnot understanding or a mind.", "relationsNarrative": "Turing-test\nThe Chinese Room Argument challenges the idea of judging intelligence by human-like chat.\n\nSymbolic AI\nThe Chinese Room Argument questions whether symbol rules can create real understanding.\n\nAI Consciousness\nThe Chinese Room Argument asks whether a machine can truly have a mind.\n\nPhysical Symbol System Hypothesis\nThe Chinese Room Argument pushes back on the claim that symbol processing is enough for intelligence.", "relations": { "turing-test": { "label": "challenges … by appearance", "note": "Sounding human does not prove understanding." }, "symbolic-ai": { "label": "questions … about understanding", "note": "Moving symbols by rules may not create meaning." }, "ai-consciousness-debate": { "label": "drives …", "note": "It asks whether a machine can truly have a mind." }, "physical-symbol-system-hypothesis": { "label": "pushes back on …", "note": "Symbol processing alone may not be intelligence." } } }, "zh": { "fullName": "中文房间论证", "factExplain": "认为符号操作不等于真正理解的思想实验。", "humanExplain": "中文房间像不会中文的网店客服：照模板回单，每句都像懂，脑子里其实没拆封。\n\n它质疑对话外观，会回复不等于理解或有意识。", "humanExplainDisplay": "中文房间像==不会中文的网店客服==：\n照模板回单，\n每句都像懂，\n脑子里其实==没拆封==。\n\n它质疑对话外观，\n会回复，\n不等于理解或有意识。", "relationsNarrative": "Turing Test\n它挑战“像人一样对话就算智能”的标准。\n\nSymbolic AI\n它质疑只靠符号规则能否产生真正理解。\n\nAI Consciousness\n它常被用来追问机器是否真的有心智。\n\nPhysical Symbol System Hypothesis\n它反驳符号处理足以构成智能的强主张。", "relations": { "turing-test": { "label": "挑战…的外观标准", "note": "像会聊，不等于真理解。" }, "symbolic-ai": { "label": "质疑…的理解观", "note": "规则搬符号，未必产生语义。" }, "ai-consciousness-debate": { "label": "支撑…的核心追问", "note": "它追问机器是否真的有心智。" }, "physical-symbol-system-hypothesis": { "label": "反驳…的强主张", "note": "符号处理未必就是智能。" } } } } }, { "id": "cifar-10", "name": "CIFAR-10", "layer": "L5", "sublayer": "product", "era": "2009", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "image-classification" }, { "to": "cnn" }, { "to": "mnist" }, { "to": "imagenet" } ], "track": "history", "i18n": { "en": { "fullName": "CIFAR-10 Dataset", "factExplain": "A dataset with 60,000 tiny 32×32 color images in 10 classes.", "humanExplain": "CIFAR-10 is like a shoebox of tiny photo stickers. Cats and trucks are so small, even Grandma squints.\n\nPeople use it to train image classifiers and test CNNs. It is small and fast, so models are easy to compare.", "humanExplainDisplay": "CIFAR-10 is like a ==shoebox of tiny photo stickers==.\nCats and trucks are ==so small, even Grandma squints==.\n\nPeople use it to train image classifiers\nand test CNNs.\nIt is small and fast,\nso models are easy to compare.", "relationsNarrative": "Image Class.\nCIFAR-10 turns Image Class into practice with ten tiny photo groups.\n\nCNN\nCNNs use CIFAR-10 to check vision skills fast.\n\nMNIST\nCIFAR-10 upgrades MNIST from digits to color objects.\n\nImageNet\nCIFAR-10 is a small warm-up before ImageNet.", "relations": { "image-classification": { "label": "used to train …", "note": "Ten tiny image classes make Image Class easy to practice." }, "cnn": { "label": "trains …", "note": "Researchers used it early to test CNNs quickly." }, "mnist": { "label": "steps up from …", "note": "It moves past digits into small color photos." }, "imagenet": { "label": "mini warm-up for …", "note": "It works like a tiny practice book before ImageNet." } } }, "zh": { "fullName": "十类小图像数据集", "factExplain": "含6万张32×32彩色图像的10类数据集。", "humanExplain": "CIFAR-10像十元店盲盒架：车猫狗船挤小格，糊到亲妈也眯眼。\n\n用于图像分类和CNN评测，小而快，便于对比模型。", "humanExplainDisplay": "CIFAR-10像十元店\n==盲盒架==：\n车猫狗船挤小格，\n糊到==亲妈也眯眼==。\n\n用于图像分类和CNN评测，\n小而快，\n便于对比模型。", "relationsNarrative": "Image Classification\n它把图像分类变成十类小图的标准练习。\n\nCNN\nCNN 常用它快速验证视觉模型好不好使。\n\nMNIST\n它比 MNIST 更难，从数字升级到彩色物体。\n\nImageNet\n它规模小得多，常作为 ImageNet 前的热身题。", "relations": { "image-classification": { "label": "用于…训练", "note": "十类小图让分类任务很直观。" }, "cnn": { "label": "训练…", "note": "早期常拿它验证卷积网络效果。" }, "mnist": { "label": "接棒…", "note": "比手写数字更接近真实照片。" }, "imagenet": { "label": "小型对照…", "note": "它像 ImageNet 的迷你练习册。" } } } } }, { "id": "classification", "name": "Classification", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "cross-entropy-loss" }, { "to": "softmax" }, { "to": "cnn" } ], "track": "history", "i18n": { "en": { "fullName": "Classification Task", "factExplain": "A learning task that puts an input into one of several preset categories.", "humanExplain": "Classification is like sorting laundry into labeled baskets. Even the weird shirt must pick a basket.\n\nIt puts each input into a fixed category. You see it in photo labels and spam filters. It also checks text.", "humanExplainDisplay": "Classification is like sorting laundry\ninto ==labeled baskets==.\nEven the ==weird shirt==\nmust pick a basket.\n\nIt puts each input\ninto a fixed category.\nYou see it in photo labels\nand spam filters.\nIt also checks text.", "relationsNarrative": "Supervised Learning\nClassification is one of the most common tasks in Supervised Learning.\n\nCross-Entropy Loss\nCross-Entropy Loss measures the gap between the prediction and the true label.\n\nSoftmax\nSoftmax turns model scores into probabilities for each category.\n\nCNN\nImage classification helped make CNNs popular.", "relations": { "supervised-learning": { "label": "often serves as a … task", "note": "Classification is one of the most common supervised learning tasks." }, "cross-entropy-loss": { "label": "often trains with …", "note": "Cross-Entropy Loss measures how wrong a class prediction is." }, "softmax": { "label": "often outputs with …", "note": "Softmax turns class scores into probabilities you can compare." }, "cnn": { "label": "helped popularize …", "note": "Image classification was a big early win for CNNs." } } }, "zh": { "fullName": "分类任务", "factExplain": "把输入判定到预设类别中的学习任务。", "humanExplain": "分类就像中医把脉后先辨证：你到底是风寒、风热，还是单纯熬夜上火，不能乱开方。\n\n它把输入归到既定类别，常见于识图、垃圾邮件过滤和文本判断。", "humanExplainDisplay": "分类就像中医把脉后\n先==辨证==：\n你到底是风寒、风热，\n还是单纯熬夜上火，\n不能==乱开方==。\n\n它把输入归到既定类别，\n常见于识图、垃圾邮件过滤\n和文本判断。", "relationsNarrative": "Supervised Learning\n分类是监督学习中最典型、最常见的任务之一。\n\nCross-Entropy Loss\n分类模型训练时，常用它来衡量预测和真值差距。\n\nSoftmax\n它常把模型输出分数变成各类别的概率分布。\n\nCNN\n图像分类的成功，曾大大推动 CNN 的普及。", "relations": { "supervised-learning": { "label": "常作为…任务", "note": "分类是监督学习里最常见的任务之一。" }, "cross-entropy-loss": { "label": "常配合…训练", "note": "多分类训练时常用它来衡量预测误差。" }, "softmax": { "label": "常用…出结果", "note": "它把各类别分数转成可比较的概率。" }, "cnn": { "label": "曾推动…落地", "note": "图像分类曾是 CNN 爆发的代表场景。" } } } } }, { "id": "classifier-free-guidance", "name": "CFG", "layer": "L2", "era": "2021", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "prompt" }, { "to": "text-to-image-generation" } ], "track": "history", "i18n": { "en": { "fullName": "Classifier-Free Guidance", "factExplain": "A sampling setting that controls how strongly the model follows the prompt.", "humanExplain": "CFG is like a volume knob for a bossy art teacher. Turn it too high, and every sunset starts yelling in neon.\n\nIn image AI, it controls how hard the picture follows your prompt. Too much can make stiff or strange images.", "humanExplainDisplay": "CFG is like a ==volume knob==\nfor a ==bossy art teacher==.\nTurn it too high,\nand every sunset starts yelling in neon.\n\nIn image AI,\nit controls how hard the picture follows your prompt.\nToo much can make stiff or strange images.", "relationsNarrative": "Diffusion\nCFG guides Diffusion sampling toward the prompt.\n\nPrompt\nThe Prompt gives CFG the target to push toward.\n\nText-to-Image Generation\nText-to-Image Generation uses CFG to balance prompt fit and a natural look.", "relations": { "diffusion": { "label": "guides … sampling", "note": "CFG pushes diffusion sampling harder toward the prompt." }, "prompt": { "label": "boosts … influence", "note": "A clear prompt gives CFG a direction to push." }, "text-to-image-generation": { "label": "tunes … results", "note": "Text-to-image uses CFG to balance prompt fit and natural images." } } }, "zh": { "fullName": "Classifier-Free Guidance，免分类器引导", "factExplain": "在采样时调节模型服从条件强度的方法。", "humanExplain": "CFG像把修图师拉进甲方群：越催越照办，过头就蓝天发光。\n\n文生图用它控贴题度；太高会僵硬出怪图。", "humanExplainDisplay": "CFG像把修图师\n拉进==甲方群==：\n越催越照办，\n过头就==蓝天发光==。\n\n文生图用它控贴题度；\n太高会僵硬出怪图。", "relationsNarrative": "Diffusion\n它在扩散采样时，放大文本条件的方向。\n\nPrompt\n提示词提供目标，它决定要往哪边加力。\n\nText-to-Image Generation\n文生图常用它平衡贴提示和画面自然度。", "relations": { "diffusion": { "label": "引导…采样", "note": "它在扩散采样时放大提示方向。" }, "prompt": { "label": "放大…影响", "note": "提示词越明确，引导越有抓手。" }, "text-to-image-generation": { "label": "调节…成图", "note": "文生图常用它平衡贴题和自然。" } } } } }, { "id": "claude-code", "name": "Claude Code", "layer": "L5", "sublayer": "product", "era": "2025", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "claude" }, { "to": "agentic-coding" }, { "to": "computer-use" }, { "to": "cursor" } ], "track": "history", "i18n": { "en": { "fullName": "Claude Code", "factExplain": "An AI coding agent built to change code and run developer tasks.", "humanExplain": "Claude Code is like a coding buddy who borrows your keyboard. You name the bug, and it starts fixing the mess.\n\nIt handles real projects, bug hunts, and coding chores fast. Big changes still need human review.", "humanExplainDisplay": "Claude Code is like a coding buddy\nwho ==borrows your keyboard==.\nYou name the bug,\nand it starts ==fixing the mess==.\n\nIt handles real projects,\nbug hunts,\nand coding chores fast.\nBig changes still need human review.", "relationsNarrative": "Claude\nClaude Code uses Claude as its model brain for real coding work.\n\nAgentic coding\nClaude Code is agentic coding because it can suggest changes and do them.\n\nComputer use\nClaude Code uses computer use when it edits files and runs terminal commands.\n\nCursor\nClaude Code and Cursor both try to be an AI coding partner.", "relations": { "claude": { "label": "is powered by …", "note": "It puts Claude into the coding workflow." }, "agentic-coding": { "label": "shows … in action", "note": "It turns coding chat into an agent that can act." }, "computer-use": { "label": "extends into …", "note": "It can edit files and run commands in the workspace." }, "cursor": { "label": "competes with …", "note": "Both bring AI into real software work." } } }, "zh": { "fullName": "Claude 编程助手", "factExplain": "面向编程任务的 AI 代码代理产品。", "humanExplain": "像武侠里替你出手的师兄：你报个招式和目标，他就下场改代码、跑命令、收残局。\n\n常用于改项目、查 bug 和自动化开发；干活很快，但关键改动还得人把关。", "humanExplainDisplay": "像武侠里替你出手的\n==师兄==：\n你报个招式和目标，\n他就下场==改代码、跑命令==。\n\n常用于改项目、查 bug\n和自动化开发；\n干活很快，\n但关键改动还得人把关。", "relationsNarrative": "Claude\n它基于 Claude 的模型能力，面向真实编程任务封装成交付产品。\n\nAgentic coding\n它是代理式编程的代表产品，让 AI 不只提建议，还会动手执行。\n\nComputer use\n它把模型能力延伸到终端和文件操作，属于计算机操作的一种落地形态。\n\nCursor\n它与 Cursor 都想成为程序员的 AI 搭子，只是入口和交互方式不同。", "relations": { "claude": { "label": "基于…能力", "note": "它把 Claude 的能力装进编程工作流。" }, "agentic-coding": { "label": "是…典型产品", "note": "它把写代码从对话变成可执行代理。" }, "computer-use": { "label": "扩展到…操作", "note": "不只会答题，还能直接动手改环境。" }, "cursor": { "label": "与…同类竞争", "note": "两者都主打 AI 参与真实开发流程。" } } } } }, { "id": "claude-fable-5", "name": "Claude Fable 5", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "claude" }, { "to": "llm" }, { "to": "computer-use" } ], "track": "today", "i18n": { "en": { "fullName": "Claude Fable 5", "factExplain": "A model version in the Claude family.", "humanExplain": "Claude Fable 5 is the group-project friend with the instructions memorized. You mumble half an idea and it opens the doc.\n\nYou meet it in office work, like drafts and questions. Add tools, and it can handle harder computer tasks.", "humanExplainDisplay": "Claude Fable 5 is the ==group-project friend==\nwith the instructions memorized.\nYou mumble ==half an idea==\nand it opens the doc.\n\nYou meet it in office work,\nlike drafts and questions.\nAdd tools,\nand it can handle harder computer tasks.", "relationsNarrative": "Claude\nClaude Fable 5 is usually seen as a version in the Claude family.\n\nLLM\nIt is still an LLM, so it understands and writes text.\n\nComputer use\nWith tools, it can handle computer tasks on screen.", "relations": { "claude": { "label": "belongs to … family", "note": "It is usually seen as a version in the Claude family." }, "llm": { "label": "is a kind of …", "note": "It is still an LLM built for talking and text." }, "computer-use": { "label": "can expand into …", "note": "With tools, it can work on screen tasks." } } }, "zh": { "fullName": "Claude Fable 5（Claude 系列模型/版本）", "factExplain": "Claude 系列中的一代模型版本。", "humanExplain": "像工位旁那个靠谱搭子：你话说半截它也能接上，文案、回复、杂活都能顺手办。\n\n常用于写作、问答和办公，也可接工具处理更复杂任务。", "humanExplainDisplay": "像工位旁那个\n==靠谱搭子==：\n你话说半截它也能接上，\n文案、回复、杂活都能\n==顺手办==。\n\n常用于写作、问答和办公，\n也可接工具处理\n更复杂任务。", "relationsNarrative": "Claude\n它通常被视为 Claude 家族中的一个版本或成员。\n\nLLM\n本质上它仍是大语言模型，擅长理解和生成文本。\n\nComputer use\n接入工具能力后，它可进一步执行电脑操作类任务。", "relations": { "claude": { "label": "属于…系列", "note": "它通常被看作 Claude 家族中的版本。" }, "llm": { "label": "是一种…", "note": "本质上仍是可对话的大语言模型。" }, "computer-use": { "label": "可扩展到…", "note": "接入工具后能处理屏幕与操作任务。" } } } } }, { "id": "claude-science", "name": "Claude Science", "layer": "L5", "sublayer": "product", "era": "2026", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "claude" }, { "to": "ai-for-science" }, { "to": "ai-assisted-research" } ], "track": "today", "i18n": { "en": { "fullName": "Claude Science", "factExplain": "A Claude-based tool that helps scientists with research work.", "humanExplain": "Claude Science is like a lab intern with bottomless coffee. It reads papers and never steals first author.\n\nYou meet it in reviews, lab plans, and data work. It clears the chores before real science starts.", "humanExplainDisplay": "Claude Science is like a lab intern\nwith ==bottomless coffee==.\nIt reads papers\nand never steals ==first author==.\n\nYou meet it in reviews,\nlab plans,\nand data work.\nIt clears the chores\nbefore real science starts.", "relationsNarrative": "Claude\nClaude is the base model that understands and writes for it.\n\nAI for Science\nClaude Science shows AI for Science as a real product.\n\nAI-assisted Research\nClaude Science brings papers, experiments, and data into daily research.", "relations": { "claude": { "label": "built on …", "note": "Claude is the base model behind this research helper." }, "ai-for-science": { "label": "turns … into a product", "note": "It is AI for Science made into a usable tool." }, "ai-assisted-research": { "label": "boosts …", "note": "It helps with reading papers and analyzing data." } } }, "zh": { "fullName": "Claude 科研助手", "factExplain": "基于 Claude 的科学研究辅助产品。", "humanExplain": "Claude Science 像实验室不睡觉的师兄：文献啃得动，数据算得快，还不抢一作。\n\n用于综述、实验设计和数据分析，先替科研清杂活。", "humanExplainDisplay": "Claude Science 像实验室==不睡觉的师兄==：\n文献啃得动、\n数据算得快，\n还==不抢一作==。\n\n用于综述、实验设计和数据分析，\n先替科研清杂活。", "relationsNarrative": "Claude\nClaude 是它的底座模型，负责理解与生成。\n\nAI for Science\n它是科学智能从概念走向产品的例子。\n\nAI-assisted Research\n它把文献、实验与数据分析接进日常研究。", "relations": { "claude": { "label": "基于…定制", "note": "它把底座模型变成科研助手。" }, "ai-for-science": { "label": "落地…", "note": "它是科学智能的一种产品化。" }, "ai-assisted-research": { "label": "强化…", "note": "它覆盖读文献和做分析。" } } } } }, { "id": "claude-sonnet-5", "name": "Claude Sonnet 5", "layer": "L3", "era": "2026", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "claude" }, { "to": "llm" }, { "to": "frontier-model" }, { "to": "claude-code" } ], "track": "today", "i18n": { "en": { "fullName": "Anthropic Claude Sonnet 5 General Model", "factExplain": "A general-purpose large language model from Anthropic’s Claude family.", "humanExplain": "Sonnet 5 is like the steady office helper. It drafts the report, and the bill does not make the boss spill coffee.\n\nPeople use it for writing and coding. It can also power agent tasks. It tries to balance ability, speed, and cost.", "humanExplainDisplay": "Sonnet 5 is like the ==steady office helper==.\nIt drafts the report,\nand the bill does not make\n==the boss spill coffee==.\n\nPeople use it for writing and coding.\nIt can also power agent tasks.\nIt tries to balance ability, speed, and cost.", "relationsNarrative": "Claude\nSonnet 5 is a Sonnet-tier model in the Claude family.\n\nLLM\nSonnet 5 is a large model for general language tasks.\n\nFrontier model\nSonnet 5 sits near the high-skill commercial model tier.\n\nClaude Code\nClaude Code can call Sonnet 5 for coding tasks.", "relations": { "claude": { "label": "belongs to …", "note": "It is part of the Claude model family." }, "llm": { "label": "is a kind of …", "note": "Its core skill comes from a large language model." }, "frontier-model": { "label": "sits near …", "note": "It is positioned near the high-skill frontier model tier." }, "claude-code": { "label": "can power …", "note": "Claude Code can use it to write and fix code." } } }, "zh": { "fullName": "Anthropic Claude Sonnet 5 通用模型", "factExplain": "Anthropic 的通用大语言模型之一。", "humanExplain": "Sonnet 5 像公司里会写方案的二把手：活接得住，账单也不吓老板。\n\n用于写作、编程和代理任务，主打能力、速度、成本平衡。", "humanExplainDisplay": "Sonnet 5 像公司里\n会写方案的==二把手==：\n活接得住，\n账单也==不吓老板==。\n\n用于写作、编程和代理任务，\n主打能力、速度、成本平衡。", "relationsNarrative": "Claude\n它是 Claude 系列中的 Sonnet 档模型。\n\nLLM\n它本质上是面向通用语言任务的大模型。\n\nFrontier model\n它代表商业前沿模型的一支。\n\nClaude Code\nClaude Code 可调用它完成编程任务。", "relations": { "claude": { "label": "属于…", "note": "它是 Claude 系列的一员。" }, "llm": { "label": "是一种…", "note": "核心能力来自大语言模型。" }, "frontier-model": { "label": "接近…", "note": "定位在高能力前沿模型梯队。" }, "claude-code": { "label": "可驱动…", "note": "常被用于代码生成与修复。" } } } } }, { "id": "claude", "name": "Claude", "layer": "L3", "era": "2023", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "chatgpt" }, { "to": "api" }, { "to": "frontier-model" } ], "track": "history", "i18n": { "en": { "fullName": "Claude Large Language Model", "factExplain": "A general-purpose LLM family made by Anthropic.", "humanExplain": "Claude is the quiet kid in the group project. It brings the neat binder and reads the instructions.\n\nPeople use it for long writing. They also use it to read documents and help with code. It fits teams that want a steady AI assistant.", "humanExplainDisplay": "Claude is the ==quiet kid==\nin the group project.\nIt brings the ==neat binder==\nand reads the instructions.\n\nPeople use it for long writing.\nThey also use it to read documents\nand help with code.\nIt fits teams that want\na steady AI assistant.", "relationsNarrative": "LLM\nClaude belongs to the LLM family.\n\nChatGPT\nPeople often compare Claude with ChatGPT.\n\nAPI\nDevelopers call Claude through an API.\n\nFrontier model\nClaude is often grouped with frontier models.", "relations": { "llm": { "label": "is a kind of …", "note": "Claude is a general LLM for chat and writing." }, "chatgpt": { "label": "is often compared with …", "note": "Both are well-known AI assistants." }, "api": { "label": "connects through an …", "note": "Developers use the API to add Claude to apps." }, "frontier-model": { "label": "belongs with …", "note": "Claude is often seen as a leading frontier model." } } }, "zh": { "fullName": "Claude 大语言模型", "factExplain": "Anthropic 推出的通用大语言模型系列。", "humanExplain": "Claude 有点像相亲局里那个话少但靠谱的对象：不负责整活抢镜，聊久了你会觉得这人能处。\n\n它常用于长文写作、文档分析和代码协作，适合稳健的企业场景。", "humanExplainDisplay": "Claude 有点像相亲局里\n那个==话少但靠谱==的对象：\n不负责整活抢镜，\n聊久了你会觉得\n这人==能处==。\n\n它常用于长文写作、\n文档分析和代码协作，\n适合稳健的企业场景。", "relationsNarrative": "LLM\n它本质上就是可对话、可生成的大语言模型。\n\nChatGPT\n它常与 ChatGPT 一起被比较体验、风格与能力。\n\nAPI\n开发者通常通过 API 调用它，接入应用和工作流。\n\nFrontier model\n它通常被归入当前能力最强的一批前沿模型。", "relations": { "llm": { "label": "属于…一类", "note": "它本质上就是一类通用大模型。" }, "chatgpt": { "label": "常被拿来对比", "note": "两者都是主流 AI 助手代表。" }, "api": { "label": "可通过…接入", "note": "开发者常用接口把它接进产品。" }, "frontier-model": { "label": "属于…阵营", "note": "它常被视作前沿模型代表之一。" } } } } }, { "id": "clip", "name": "CLIP", "layer": "L3", "era": "2021", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "embedding" }, { "to": "diffusion" }, { "to": "pretraining" } ], "track": "history", "i18n": { "en": { "fullName": "Contrastive Language-Image Pre-training", "factExplain": "A multimodal model that maps images and text into one shared meaning space.", "humanExplain": "CLIP is like the name-tag helper on school picture day. It knows “golden retriever” matches the dog photo, not your math teacher.\n\nIt can match images with captions and find pictures from words. It also helps some text-to-image AIs read prompts.", "humanExplainDisplay": "CLIP is like the ==name-tag helper==\non school picture day.\nIt knows “golden retriever” ==matches the dog photo==,\nnot your math teacher.\n\nIt can match images with captions\nand find pictures from words.\nIt also helps some text-to-image AIs\nread prompts.", "relationsNarrative": "Multimodal AI\nCLIP is a classic way to match pictures with words.\n\nEmbedding\nCLIP turns images and text into number lists you can compare.\n\nDiffusion\nMany text-to-image systems use CLIP to understand prompts better.\n\nPretraining\nCLIP learns from huge sets of image-caption pairs first.", "relations": { "multimodal": { "label": "shows … at work", "note": "CLIP is a classic way to match images with language." }, "embedding": { "label": "creates … representations", "note": "It turns images and text into number lists you can compare." }, "diffusion": { "label": "helps … read prompts", "note": "Many text-to-image tools use it to match prompts with images." }, "pretraining": { "label": "learns matching through …", "note": "It first studies huge sets of image-caption pairs." } } }, "zh": { "fullName": "对比语言-图像预训练", "factExplain": "把图片和文字映射到同一语义空间的多模态模型。", "humanExplain": "给它一张图再配一句话，它就像活动现场的对号员，扫两眼就知道这俩是不是一伙的。\n\n能做图文匹配和以文搜图，也常帮文生图模型理解提示词。", "humanExplainDisplay": "给它一张图再配一句话，\n它就像活动现场的==对号员==，\n扫两眼就知道这俩\n是不是==一伙的==。\n\n能做图文匹配和以文搜图，\n也常帮文生图模型理解提示词。", "relationsNarrative": "Multimodal AI\n它是多模态里最经典的图文对齐方法之一。\n\nEmbedding\n它会把图片和文字变成可比较的向量表示。\n\nDiffusion\n很多文生图系统会借它提升对提示词的理解。\n\nPretraining\n它依赖海量图文配对数据做预训练。", "relations": { "multimodal": { "label": "属于…基础能力", "note": "它是图像和语言对齐的经典方法。" }, "embedding": { "label": "生成…表示", "note": "它把图文都压成可比较向量。" }, "diffusion": { "label": "帮助…懂提示词", "note": "不少文生图系统借它做图文对齐。" }, "pretraining": { "label": "通过…学对齐", "note": "它先在海量图文对上预训练。" } } } } }, { "id": "closed-model", "name": "Closed-source Model", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "open-weights" }, { "to": "api" }, { "to": "maas-model-as-a-service" } ], "track": "today", "i18n": { "en": { "fullName": "Closed-source Model", "factExplain": "A model whose weights stay private, so you usually use it through a service.", "humanExplain": "A closed model is like a burger place with secret sauce. You can buy lunch, but the cook will not show the recipe.\n\nIt helps you use a strong model fast. But checking it, running it yourself, and changing it are hard.", "humanExplainDisplay": "A closed model is like a burger place\nwith ==secret sauce==.\nYou can buy lunch,\nbut the cook ==will not show the recipe==.\n\nIt helps you use a strong model fast.\nBut checking it,\nrunning it yourself,\nand changing it are hard.", "relationsNarrative": "Open-source-model\nA closed model is the opposite of an open-source model.\n\nOpen weights\nA closed model usually keeps its weights private.\n\nAPI\nMany closed models offer their skills through an API.\n\nMaaS\nMaaS packages a closed model as a paid service.", "relations": { "open-source-model": { "label": "is the opposite of …", "note": "Openness decides if you can run or change the model yourself." }, "open-weights": { "label": "keeps … private", "note": "Private weights are the main wall around a closed model." }, "api": { "label": "is often called through …", "note": "Most closed models let outsiders use them through an API." }, "maas-model-as-a-service": { "label": "is delivered as …", "note": "MaaS turns the model into a paid service you can use on demand." } } }, "zh": { "fullName": "闭源模型", "factExplain": "不公开权重，通常只提供调用的模型。", "humanExplain": "闭源模型像奶茶店秘方：你能点单喝到爽，珍珠怎么煮老板不让看。\n\n适合快速接入强模型，但审计、自部署和改造受限。", "humanExplainDisplay": "闭源模型像奶茶店秘方：\n你能==点单喝到爽==，\n珍珠怎么煮，\n==老板不让看==。\n\n适合快速接入强模型，\n但审计、自部署，\n和改造受限。", "relationsNarrative": "Open-source Model\n它与闭源模型相反，强调代码或权重可获取。\n\nOpen Weights\n闭源模型通常不放出权重，只让你远程使用。\n\nAPI\n许多闭源模型通过 API 对外提供能力。\n\nMaaS\nMaaS 把闭源模型包装成可计费的服务。", "relations": { "open-source-model": { "label": "与…相反", "note": "开放程度决定能否自部署和改造。" }, "open-weights": { "label": "不开放…", "note": "权重不公开，是封闭的核心边界。" }, "api": { "label": "常通过…调用", "note": "多数闭源模型靠 API 让外部调用。" }, "maas-model-as-a-service": { "label": "以…交付", "note": "服务化交付让模型像水电一样用。" } } } } }, { "id": "clustering", "name": "Clustering", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "unsupervised-learning" }, { "to": "classification" }, { "to": "gaussian-mixture-model" }, { "to": "dimensionality-reduction" } ], "track": "history", "i18n": { "en": { "fullName": "Clustering", "factExplain": "An unsupervised method that automatically groups data by similarity.", "humanExplain": "Clustering is like a messy sock drawer sorting itself. Striped socks find striped socks. Fuzzy socks form a tiny club.\n\nIt groups messy data before anyone names the groups. You meet it in customer groups, weird-point alerts, and photo sorting.", "humanExplainDisplay": "Clustering is like a messy sock drawer ==sorting itself==.\nStriped socks find striped socks.\nFuzzy socks form a ==tiny club==.\n\nIt groups messy data before anyone names the groups.\nYou meet it in customer groups,\nweird-point alerts,\nand photo sorting.", "relationsNarrative": "Unsupervised Learning\nClustering is a classic Unsupervised Learning task with no labels.\n\nClassification\nClassification uses known labels, but clustering makes groups first.\n\nGMM\nGMM is a common probability-based way to do clustering.\n\nDim. Reduction\nDim. Reduction can spread data out before clustering.", "relations": { "unsupervised-learning": { "label": "is a kind of …", "note": "Clustering is a classic task with no human labels." }, "classification": { "label": "contrasts with …", "note": "Classification uses known labels. Clustering makes groups first." }, "gaussian-mixture-model": { "label": "can use …", "note": "GMM is a common probability-based way to cluster data." }, "dimensionality-reduction": { "label": "often works with …", "note": "Dim. Reduction can spread data out before clustering." } } }, "zh": { "fullName": "聚类", "factExplain": "按相似性自动把数据分成若干组的无监督方法。", "humanExplain": "食堂下课高峰抢座，不用贴标签，情侣会挨情侣，社牛会拼社牛，大家自己就坐成几拨。\n\n它用于用户分群、异常发现和图像整理，先把乱数据分出圈层。", "humanExplainDisplay": "食堂下课高峰抢座，\n不用贴标签，\n情侣会挨情侣，\n社牛会拼==社牛==，\n大家自己就坐成==几拨==。\n\n它用于用户分群、异常发现\n和图像整理，\n先把乱数据分出圈层。", "relationsNarrative": "Unsupervised Learning\n聚类是不依赖标签的典型无监督学习任务。\n\nClassification\n分类按已知标签判断，聚类则先自己分组。\n\nGaussian Mixture Model\nGMM 是一种常见的概率式聚类实现方法。\n\nDimensionality Reduction\n降维常先把数据摊开，方便聚类和观察结构。", "relations": { "unsupervised-learning": { "label": "属于…方法", "note": "它是不靠人工标签的典型任务。" }, "classification": { "label": "常与…对比", "note": "一个是自动分堆，一个是按标签判断。" }, "gaussian-mixture-model": { "label": "可用…实现", "note": "GMM 是常见的概率式聚类方法。" }, "dimensionality-reduction": { "label": "常配合…使用", "note": "先降维后聚类更便于分组与可视化。" } } } } }, { "id": "cnn", "name": "CNN", "layer": "L3", "era": "1989", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "alexnet" }, { "to": "resnet" }, { "to": "transformer" } ], "track": "history", "i18n": { "en": { "fullName": "Convolutional Neural Network", "factExplain": "A neural network built to understand image-like grids.", "humanExplain": "CNN is like a kid solving a hidden-picture puzzle. It spots little edges first, then yells, “cat!”\n\nIt helps sort and recognize images. You meet it in photo apps and video tools.", "humanExplainDisplay": "CNN is like a kid solving a ==hidden-picture puzzle==.\nIt spots ==little edges== first,\nthen yells, “cat!”\n\nIt helps sort and recognize images.\nYou meet it in photo apps and video tools.", "relationsNarrative": "Neural-network\nCNN is a classic vision branch of the neural network family.\n\nAlexNet\nAlexNet made CNNs a main tool for computer vision again.\n\nResNet\nResNet is a famous upgrade from deeper, better CNNs.\n\nTransformer\nCNN is often compared with Transformer for different task strengths.", "relations": { "neural-network": { "label": "is a kind of …", "note": "CNN is a classic vision branch of neural networks." }, "alexnet": { "label": "was boosted by …", "note": "AlexNet made CNNs famous in computer vision." }, "resnet": { "label": "grew into …", "note": "ResNet is a stronger, deeper version of the CNN line." }, "transformer": { "label": "is often compared with …", "note": "Both are key designs, but they shine at different jobs." } } }, "zh": { "fullName": "卷积神经网络", "factExplain": "一种擅长处理图像网格数据的神经网络。", "humanExplain": "它像老中医看舌苔：先瞄纹路颜色，再综合判断，最后分出这是上火还是没睡够。\n\n常做图像分类和识别，适合处理照片、视频等视觉任务。", "humanExplainDisplay": "它像老中医看舌苔：\n先瞄==纹路颜色==，\n再综合判断，\n最后分出==这是上火还是没睡够==。\n\n常做图像分类和识别，\n适合处理照片、\n视频等视觉任务。", "relationsNarrative": "Neural-network\n它是神经网络家族里最经典的视觉分支之一。\n\nAlexNet\nAlexNet 的成功让它重新成为计算机视觉主流。\n\nResNet\nResNet 是它不断加深、改进后的代表性演进。\n\nTransformer\n它常被拿来和 Transformer 对比不同任务优势。", "relations": { "neural-network": { "label": "属于…一类", "note": "它是神经网络在视觉上的经典分支。" }, "alexnet": { "label": "被…带火", "note": "AlexNet 让它在视觉领域一战成名。" }, "resnet": { "label": "发展到…", "note": "ResNet 是这条路线上的代表升级版。" }, "transformer": { "label": "常与…对比", "note": "二者都是重要架构，擅长方向不同。" } } } } }, { "id": "coco-dataset", "name": "COCO Dataset", "layer": "L5", "sublayer": "product", "era": "2014", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "imagenet" }, { "to": "cnn" }, { "to": "clip" } ], "track": "history", "seo": { "en": { "title": "What Is the COCO Dataset? The Driving Test for Vision Models", "description": "COCO benchmarks object detection, segmentation, and captioning — not just what's in the image, but exactly where. A plain-English explainer with related concepts." }, "zh": { "title": "COCO 数据集是什么?视觉模型的驾照路考,一文看懂 — AI Rookies", "description": "认出是人是车还不够,还得标清在哪、边界到哪。目标检测、分割、图像描述的标配数据集,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "COCO Dataset (Common Objects in Context)", "factExplain": "A vision dataset for object detection, image segmentation, and image captioning.", "humanExplain": "COCO is like a messy birthday photo covered in bossy sticky notes. The notes box the cake, trace the dog, and caption the frosting attack.\n\nIt trains and tests vision models. You meet it in object detection, segmentation, and image captioning.", "humanExplainDisplay": "COCO is like a messy birthday photo\ncovered in ==bossy sticky notes==.\nThe notes box the cake,\ntrace the dog,\nand ==caption the frosting attack==.\n\nIt trains and tests vision models.\nYou meet it in object detection,\nsegmentation,\nand image captioning.", "relationsNarrative": "Computer Vision\nCOCO is one of the most common benchmark datasets in Computer Vision.\n\nImageNet\nBoth are classic vision datasets, but COCO focuses more on location and scenes.\n\nCNN\nMany CNN models used COCO to train and compare detection and segmentation skill.\n\nCLIP\nCOCO has image captions, so it can help CLIP learn image-text matching.", "relations": { "computer-vision": { "label": "benchmarks … tasks", "note": "COCO is a common benchmark for Computer Vision tasks." }, "imagenet": { "label": "sits beside …", "note": "Both are classic vision datasets, but COCO cares more about location and scenes." }, "cnn": { "label": "trains and tests …", "note": "Many CNN models used COCO to compare detection and segmentation skill." }, "clip": { "label": "supports … image-text learning", "note": "COCO has captions, so it helps train image-text matching." } } }, "zh": { "fullName": "COCO 数据集（Common Objects in Context）", "factExplain": "一个用于目标检测、分割和图像描述的视觉数据集。", "humanExplain": "它像驾校路考表：不光要认出前面是人是车，还得标清在哪、边界到哪，顺带说清现场情况。\n\n常拿来训练和评测视觉模型，尤其用于检测、分割和图像描述。", "humanExplainDisplay": "它像驾校\n==路考表==：\n不光要认出前面\n是人是车，\n还得标清在哪、\n边界==到哪==，\n顺带说清现场情况。\n\n常拿来训练和评测\n视觉模型，\n尤其用于检测、分割\n和图像描述。", "relationsNarrative": "Computer Vision\n它是计算机视觉里最常用的基准数据集之一。\n\nImageNet\n两者都是经典视觉数据集，但它更强调定位与场景。\n\nCNN\n很多卷积模型曾用它训练或比较检测分割能力。\n\nCLIP\n它含图文标注，也常被用于图文对齐相关训练。", "relations": { "computer-vision": { "label": "服务…任务", "note": "它是视觉任务常用基准数据集。" }, "imagenet": { "label": "和…并列常见", "note": "两者都是经典视觉数据集，但侧重点不同。" }, "cnn": { "label": "供…训练评测", "note": "早期大量视觉模型在它上面比成绩。" }, "clip": { "label": "支撑…图文学习", "note": "它含图文标注，可用于图文对齐任务。" } } } } }, { "id": "codex", "name": "Codex", "layer": "L3", "era": "2021", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "copilot" }, { "to": "agentic-coding" }, { "to": "llm" }, { "to": "cursor" } ], "track": "history", "i18n": { "en": { "fullName": "Code generation model / AI coding assistant", "factExplain": "A model tuned to read, write, and fix computer code.", "humanExplain": "Codex is like a super-fast lab partner in computer class. You share the idea, and it starts typing like the keyboard owes it money.\n\nIt writes code and finishes half-done functions. It can fix bugs too, but you check it before launch.", "humanExplainDisplay": "Codex is like a ==super-fast lab partner==\nin computer class.\nYou share the idea,\nand it starts typing like\n==the keyboard owes it money==.\n\nIt writes code and finishes half-done functions.\nIt can fix bugs too,\nbut you check it before launch.", "relationsNarrative": "Copilot\nCodex often provides the code-writing power behind tools like Copilot.\n\nAgentic coding\nCodex helps AI do more than suggest code. It can help write and change it.\n\nLLM\nCodex is a kind of LLM tuned for programming work.\n\nCursor\nModels like Codex are often built into AI coding tools like Cursor.", "relations": { "copilot": { "label": "powers …", "note": "Codex often sits under AI coding helpers like Copilot." }, "agentic-coding": { "label": "supports …", "note": "Codex helps AI move from giving tips to doing code tasks." }, "llm": { "label": "is a branch of …", "note": "Codex is basically an LLM tuned for programming." }, "cursor": { "label": "is built into …", "note": "Models like Codex are often added to coding tools like Cursor." } } }, "zh": { "fullName": "代码生成模型 / AI 编程助手", "factExplain": "面向编程任务优化的代码生成与理解模型。", "humanExplain": "你刚把需求甩过去，它就跟键盘侠附体似的，哐哐先把代码敲出个八九不离十。\n\n常用来写代码、补全函数和改 Bug，但上线前仍得人自己检查。", "humanExplainDisplay": "你刚把需求甩过去，\n它就跟==键盘侠附体==似的，\n哐哐先把代码敲出个\n==八九不离十==。\n\n常用来写代码、\n补全函数和改 Bug，\n但上线前仍得人自己检查。", "relationsNarrative": "Copilot\n它常作为 Copilot 这类编程助手的底层生成能力。\n\nAgentic-coding\n它让 AI 不只提建议，还能直接参与写改代码。\n\nLLM\n它本质上是针对编程场景优化的一类大语言模型。\n\nCursor\n这类模型常被集成进 Cursor 等 AI 编程产品中。", "relations": { "copilot": { "label": "驱动…体验", "note": "它常作为 AI 编程助手的底层能力。" }, "agentic-coding": { "label": "支撑…执行", "note": "它让代码任务从建议走向代做。" }, "llm": { "label": "属于…分支", "note": "它本质上是面向编程优化的大模型。" }, "cursor": { "label": "嵌入…产品", "note": "这类模型常被集成进编程工具中。" } } } } }, { "id": "cognitive-model", "name": "Cognitive Model", "layer": "L1", "era": "1950s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "world-model" }, { "to": "belief-state" }, { "to": "agent" }, { "to": "knowledge-representation" } ], "track": "today", "i18n": { "en": { "fullName": "Cognitive Model", "factExplain": "A simple map of how people think, choose, and change their minds.", "humanExplain": "A cognitive model is like pausing a kid in a candy aisle. It asks, “Why gummy worms, not the sad apple?”\n\nIt helps explain how people judge and choose. AI builders use it when they design Agent reasoning and decisions.", "humanExplainDisplay": "A cognitive model is like pausing a kid\nin a ==candy aisle==.\nIt asks,\n“==Why gummy worms==,\nnot the sad apple?”\n\nIt helps explain how people judge and choose.\nAI builders use it when they design\nAgent reasoning and decisions.", "relationsNarrative": "World model\nA cognitive model can show how someone builds an inner view of the world.\n\nBelief State\nIn thinking, the Belief State changes as new clues arrive.\n\nAgent\nA cognitive model can guide an Agent's reasoning and choices.\n\nKR\nA cognitive model needs KR to make knowledge usable.", "relations": { "world-model": { "label": "helps explain …", "note": "It shows how someone builds an inner view of the world." }, "belief-state": { "label": "tracks changes in …", "note": "Thinking often changes what someone believes." }, "agent": { "label": "guides … design", "note": "It can guide how an Agent reasons and decides." }, "knowledge-representation": { "label": "uses … to express ideas", "note": "KR makes the model's knowledge usable." } } }, "zh": { "fullName": "认知模型", "factExplain": "对人类思考与决策过程的抽象描述。", "humanExplain": "像教练复盘一盘棋：先看你怎么想、哪步犹豫、为何改手，不再只剩一句“我当时觉得”。\n\n用于解释人的判断行为，也常启发智能体的推理和决策设计。", "humanExplainDisplay": "像教练复盘一盘棋：\n先看你怎么想、哪步犹豫、\n为何改手，不再只剩一句\n==“我当时觉得”==。\n\n用于解释人的判断行为，\n也常启发智能体的\n推理和决策设计。", "relationsNarrative": "World Model\n认知模型常用来描述主体如何形成对世界的内部理解。\n\nBelief State\n认知过程里，信念状态会随观察与推理不断更新。\n\nAgent\n它能为 Agent 的记忆、推理与决策设计提供启发。\n\nKnowledge Representation\n认知模型里的知识内容，往往要靠知识表示来落地。", "relations": { "world-model": { "label": "帮助理解…", "note": "它常被用来刻画主体如何理解世界。" }, "belief-state": { "label": "描述…变化", "note": "认知过程常体现为信念状态更新。" }, "agent": { "label": "启发…设计", "note": "可为智能体的决策流程提供参考。" }, "knowledge-representation": { "label": "依赖…表达", "note": "认知内容需要可操作的知识表示。" } } } } }, { "id": "collaborative-filtering", "name": "Collaborative Filtering", "layer": "L2", "era": "1992", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "recommender-system" }, { "to": "matrix-factorization" }, { "to": "k-nearest-neighbors" } ], "track": "history", "i18n": { "en": { "fullName": "Collaborative Filtering", "factExplain": "A recommendation method based on similar users or similar items.", "humanExplain": "Collaborative filtering is like a school cafeteria buddy. You both love curly fries, so their favorite cookie lands on your tray next.\n\nIt recommends things from similar people or similar items. You meet it in shopping apps, video feeds, and music apps. New users make it guess badly.", "humanExplainDisplay": "Collaborative filtering is like a ==school cafeteria buddy==.\nYou both love curly fries,\nso their favorite cookie ==lands on your tray== next.\n\nIt recommends things from similar people\nor similar items.\nYou meet it in shopping apps,\nvideo feeds, and music apps.\nNew users make it guess badly.", "relationsNarrative": "Recommender System\nCollaborative filtering is one classic method inside a recommender system.\n\nMatrix Factorization\nMatrix Factorization can model hidden taste patterns for collaborative filtering.\n\nKNN\nKNN can find similar users or similar items.", "relations": { "recommender-system": { "label": "powers …", "note": "Collaborative filtering is a classic way to build recommendations." }, "matrix-factorization": { "label": "can use …", "note": "Matrix Factorization turns taste patterns into hidden factors." }, "k-nearest-neighbors": { "label": "finds similar taste with …", "note": "KNN can find users or items with similar taste." } } }, "zh": { "fullName": "协同过滤", "factExplain": "根据相似用户或物品推荐内容的方法。", "humanExplain": "协同过滤像宿舍点奶茶：你和室友口味像，他常喝的新品，就先推到你手里。\n\n常用于电商、短视频和音乐，冷启动时容易猜不准。", "humanExplainDisplay": "协同过滤像宿舍点奶茶：\n你和室友==口味像==，\n他常喝的新品，\n就==先推到你手里==。\n\n常用于电商、短视频和音乐，\n冷启动时，\n容易猜不准。", "relationsNarrative": "Recommender System\n协同过滤是推荐系统里最经典的推荐方法之一。\n\nMatrix Factorization\n矩阵分解常用来实现协同过滤的潜在偏好建模。\n\nKNN\nKNN 可用于寻找相似用户或相似物品。", "relations": { "recommender-system": { "label": "支撑…", "note": "协同过滤是推荐系统的经典做法。" }, "matrix-factorization": { "label": "可用…实现", "note": "矩阵分解把偏好压成潜在因子。" }, "k-nearest-neighbors": { "label": "用…找相似", "note": "近邻方法常用来找相似用户或物品。" } } } } }, { "id": "command-line-interface", "name": "Command Line Interface", "layer": "L6", "era": "1960s", "publishedAt": "2026-06-21T13:36:46.876Z", "relations": [ { "to": "api" }, { "to": "ollama" }, { "to": "llama-cpp" }, { "to": "agentic-coding" }, { "to": "claude-code" } ], "track": "ingest", "i18n": { "en": { "fullName": "Command Line Interface", "factExplain": "A text window where you type commands to run AI tools.", "humanExplain": "A command line is a drive-thru speaker for your computer. No buttons. You bark the order, and the kitchen starts cooking.\n\nIn AI work, people use it to start models and training jobs. You meet it in tools like Ollama and Llama.cpp.", "humanExplainDisplay": "A command line is a ==drive-thru speaker== for your computer.\nNo buttons.\nYou bark the order, and the kitchen starts cooking.\n\nIn AI work, people use it to start models and training jobs.\nYou meet it in tools like Ollama and Llama.cpp.", "relations": { "api": { "label": "can call …", "note": "A CLI can send API calls for you." }, "ollama": { "label": "runs … with commands", "note": "Ollama is often used from a CLI." }, "llama-cpp": { "label": "starts … with commands", "note": "Llama.cpp is a classic CLI AI tool." }, "agentic-coding": { "label": "powers … workflows", "note": "Coding agents often run shell commands." } } }, "zh": { "fullName": "命令行界面", "factExplain": "用文本命令运行和控制程序的界面。", "humanExplain": "命令行界面像钻进 AI 后厨点菜：不看花哨菜单，直接喊师傅起锅。\n\n开发者用它训练模型、启动本地大模型、批量跑脚本。", "humanExplainDisplay": "命令行界面像==钻进 AI 后厨点菜==：\n不看花哨菜单，直接==喊师傅起锅==。\n\n开发者用它训练模型、\n启动本地大模型、\n批量跑脚本。", "relations": { "api": { "label": "与…同为操作入口", "note": "API 给程序用，命令行给人敲。" }, "ollama": { "label": "常用来操作…", "note": "Ollama 常靠命令行拉模型、跑对话。" }, "llama-cpp": { "label": "用来启动…", "note": "Llama.cpp 常用命令行加载本地模型。" }, "claude-code": { "label": "承载…", "note": "Claude Code 把代码代理放进终端。" } } } } }, { "id": "common-crawl", "name": "Common Crawl", "layer": "L5", "sublayer": "product", "era": "2008", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "llm" }, { "to": "big-data" }, { "to": "copyright" } ], "track": "history", "i18n": { "en": { "fullName": "Common Crawl", "factExplain": "A huge public web-page dataset often used to pretrain AI models.", "humanExplain": "Common Crawl is a giant Roomba for the public web. It bumps around and sweeps pages into a huge dustbin.\n\nAI teams use it to pretrain big models. It can also bring copyright fights, privacy leaks, and junk.", "humanExplainDisplay": "Common Crawl is a ==giant Roomba==\nfor the public web.\nIt bumps around\nand sweeps pages into a ==huge dustbin==.\n\nAI teams use it\nto pretrain big models.\nIt can also bring copyright fights,\nprivacy leaks, and junk.", "relationsNarrative": "Pretraining\nCommon Crawl often feeds pretraining, so models learn language patterns first.\n\nLLM\nLLMs often learn language and common facts from huge web datasets like this.\n\nBig Data\nCommon Crawl turns web scraping into downloadable and reusable Big Data.\n\nCopyright\nPutting web pages into training data can cross copyright lines.", "relations": { "pretraining": { "label": "feeds text to …", "note": "Many models use Common Crawl as raw text for pretraining." }, "llm": { "label": "supplies web text to …", "note": "LLMs learn language patterns from huge piles of web pages." }, "big-data": { "label": "shows the scale of …", "note": "Common Crawl turns web scraping into downloadable Big Data." }, "copyright": { "label": "raises … issues", "note": "Web pages in training data can cross copyright lines." } } }, "zh": { "fullName": "公共网页抓取数据集", "factExplain": "公开的大规模网页抓取数据集，常用于模型预训练。", "humanExplain": "Common Crawl 是互联网夜市大扫货：广告、菜谱、论文全装袋，AI 先混个眼熟。\n\n常作大模型预训练底料，也带来版权、隐私和脏数据。", "humanExplainDisplay": "Common Crawl 是\n==互联网夜市大扫货==：\n广告、菜谱、论文全装袋，\nAI 先==混个眼熟==。\n\n常作大模型预训练底料，\n也带来版权、隐私和脏数据。", "relationsNarrative": "Pretraining\n它常作为预训练语料，帮模型先学语言模式。\n\nLLM\nLLM 常从这类海量网页中学习语言和常识。\n\nBig Data\n它把网页抓取变成可下载、可复用的大数据。\n\nCopyright\n网页内容进入训练集，容易触碰版权边界。", "relations": { "pretraining": { "label": "给…供语料", "note": "许多模型预训练会用它做底料。" }, "llm": { "label": "为…提供网页语料", "note": "LLM 常从海量网页中学语言模式。" }, "big-data": { "label": "体现…规模", "note": "它把网页抓取变成可下载的大数据。" }, "copyright": { "label": "引发…争议", "note": "网页内容进训练集常碰到版权边界。" } } } } }, { "id": "commonsense-reasoning", "name": "Commonsense Reasoning", "layer": "L1", "era": "1950s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "reasoning-model" }, { "to": "world-model" }, { "to": "hallucination" } ], "track": "history", "i18n": { "en": { "fullName": "Commonsense Reasoning", "factExplain": "The skill of using everyday knowledge to make sensible guesses.", "humanExplain": "Commonsense reasoning is like a parent in the school pickup line. One look at a muddy kid says, “Do not let him sit on the white couch.”\n\nYou meet it when chatbots answer questions. Robots use it too. It feels more human, but odd cases can still fool it.", "humanExplainDisplay": "Commonsense reasoning is like a ==parent in the school pickup line==.\nOne look at a ==muddy kid== says,\n“Do not let him sit on the white couch.”\n\nYou meet it when chatbots answer questions.\nRobots use it too.\nIt feels more human,\nbut odd cases can still fool it.", "relationsNarrative": "Reasoning-model\nA reasoning model can handle steps, but it may still lack steady common sense.\n\nWorld model\nCommonsense reasoning needs a basic feel for objects, causes, and scenes.\n\nHallucination\nCommon sense can filter some silly answers, but it cannot stop every hallucination.", "relations": { "reasoning-model": { "label": "adds daily judgment to …", "note": "Strong step-by-step reasoning does not always mean strong common sense." }, "world-model": { "label": "depends on …", "note": "A better world model makes common sense feel smoother." }, "hallucination": { "label": "helps reduce …", "note": "Common sense can catch some wild answers." } } }, "zh": { "fullName": "Commonsense Reasoning｜常识推理", "factExplain": "利用日常常识进行判断与推断的能力。", "humanExplain": "它像菜市场老摊主，瞄你两眼就知道你是随便逛逛，还是今晚真要开火做饭。\n\n常见于对话、问答和机器人判断，让回答更像真人；遇到冷门情境仍可能失准。", "humanExplainDisplay": "它像==菜市场==\n老摊主，瞄你两眼\n就知道你是随便逛逛，\n还是今晚真要==开火==做饭。\n\n常见于对话、问答\n和机器人判断，\n让回答更像真人；\n遇冷门情境仍会失准。", "relationsNarrative": "Reasoning-model\n推理模型擅长算步骤，但未必天然具备稳定常识。\n\nWorld-model\n常识推理常依赖对物体、因果和场景的世界理解。\n\nHallucination\n常识能过滤部分离谱答案，但挡不住所有幻觉。", "relations": { "reasoning-model": { "label": "补足…日常判断", "note": "推理强不等于常识一定稳。" }, "world-model": { "label": "依赖…理解世界", "note": "世界模型越好，常识判断越顺。" }, "hallucination": { "label": "帮助减少…乱编", "note": "常识能拦下一部分离谱回答。" } } } } }, { "id": "companion-robot", "name": "Companion Robot", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "affective-computing" }, { "to": "ai-companion-risk" }, { "to": "digital-human" } ], "track": "today", "i18n": { "en": { "fullName": "Companion Robot", "factExplain": "A robot-shaped AI product built for company, interaction, and care.", "humanExplain": "A companion robot is like a tiny robot roommate. It chats with you and never steals your fries.\n\nIt stays close for company and care. You may meet one with older people, kids, or families at home.", "humanExplainDisplay": "A companion robot is like a ==tiny robot roommate==.\nIt chats with you\nand ==never steals your fries==.\n\nIt stays close for company and care.\nYou may meet one with older people,\nkids,\nor families at home.", "relationsNarrative": "Embodied AI\nA companion robot is a home-friendly form of Embodied AI.\n\nAffective Computing\nAffective Computing helps it notice moods and offer comfort.\n\nCompanion-risk\nThe more real the bond feels, the more dependence and false trust can grow.\n\nDigital human\nA Digital human lives on a screen. A companion robot stays beside you in the room.", "relations": { "embodied-ai": { "label": "is a real-world form of …", "note": "It puts AI into a body people can touch and talk to." }, "affective-computing": { "label": "reads feelings with …", "note": "Emotion sensing helps it respond more like a caring person." }, "ai-companion-risk": { "label": "can create …", "note": "The more real it feels, the easier dependence and false trust grow." }, "digital-human": { "label": "feels close to …", "note": "Both try to feel human and keep you company." } } }, "zh": { "fullName": "陪伴机器人", "factExplain": "以陪伴、互动和照护为目标的机器人形态 AI 产品。", "humanExplain": "有点像家里多了个不睡觉的电子室友，会聊天、会回应，还总在你身边晃，让屋里没那么空。\n\n常见于养老、儿童互动和居家照护，核心是长期陪伴与照应。", "humanExplainDisplay": "有点像家里多了个\n==不睡觉的电子室友==，\n会聊天、会回应，\n还总在你身边晃，\n让屋里==没那么空==。\n\n常见于养老、\n儿童互动和居家照护，\n核心是长期陪伴与照应。", "relationsNarrative": "Embodied AI\n它是具身智能的一种消费级落地形态。\n\nAffective Computing\n情绪识别让它更会察言观色和安抚人。\n\nCompanion-risk\n陪伴越拟真，越可能引发依赖和误判。\n\nDigital human\n数字人偏屏幕里，它偏现实中陪在身边。", "relations": { "embodied-ai": { "label": "属于…落地形态", "note": "它把智能能力装进可互动实体。" }, "affective-computing": { "label": "靠…读情绪", "note": "情绪识别让陪伴更像在交流。" }, "ai-companion-risk": { "label": "会带来…问题", "note": "陪伴越深，依赖与误导风险越高。" }, "digital-human": { "label": "和…形态相近", "note": "两者都主打拟人互动与陪伴感。" } } } } }, { "id": "compute-race", "name": "Compute-race", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T11:20:00Z", "relations": [ { "to": "gpu" }, { "to": "scaling-law" }, { "to": "foundation-model" }, { "to": "pretraining" } ], "i18n": { "en": { "fullName": "Compute Race", "factExplain": "A race between companies and countries for AI computing power.", "humanExplain": "It is a gold rush for AI. But the shovels are GPUs, and the electric bill bites.\n\nYou meet it in big model training and cloud AI prices. It makes chips strategic, and raises the entry fee for AI.", "humanExplainDisplay": "It is a ==gold rush== for AI.\nBut the shovels are ==GPUs==,\nand the electric bill bites.\n\nYou meet it in big model training\nand cloud AI prices.\nIt makes chips strategic,\nand raises the entry fee for AI.", "relationsNarrative": "GPU\nGPUs are the rare gear everyone wants in the compute race.\n\nScaling-law\nScaling-law keeps pushing teams to buy more compute.\n\nFoundation-model\nBigger Foundation-models make the compute race harder to cool down.\n\nPretraining\nPretraining uses huge compute, so it feeds the compute race.", "relations": { "gpu": { "label": "fights over …", "note": "GPUs are the scarce fuel in the compute race." }, "scaling-law": { "label": "is driven by …", "note": "Scaling-law rewards more compute, so the race keeps going." }, "foundation-model": { "label": "heats up for …", "note": "Bigger Foundation-models need more compute to train." }, "pretraining": { "label": "centers on …", "note": "Pretraining burns huge compute before a model is useful." } } }, "zh": { "fullName": "算力竞赛", "factExplain": "企业和国家围绕 AI 算力资源展开的竞争。", "humanExplain": "算力竞赛像大厂版抢春运票：谁先抢到 GPU、电和机房，谁就先发车。\n\n它影响模型上限、成本和能源需求，常见于大模型公司与国家竞争。", "humanExplainDisplay": "算力竞赛像==大厂版抢春运票==：\n谁先抢到 GPU、电和机房，\n谁就==先发车==。\n\n它影响模型上限、成本和能源需求，\n常见于大模型公司与国家竞争。", "relationsNarrative": "GPU\nGPU 是 Compute-race 中最稀缺也最关键的资源。\n\nScaling-law\nScaling-law 推动了对算力的持续竞争，最终形成 Compute-race。\n\nFoundation-model\nFoundation-model 的训练规模越大，Compute-race 越难降温。\n\nPretraining\nPretraining 持续消耗算力，是 Compute-race 的直接来源。", "relations": { "gpu": { "label": "争夺…" }, "scaling-law": { "label": "被…驱动" }, "foundation-model": { "label": "为…而起" }, "pretraining": { "label": "集中于…" } } } } }, { "id": "computer-use", "name": "Computer use", "layer": "L4", "era": "2024", "publishedAt": "2026-05-28T15:58:23.419Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "copilot" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "Computer use", "factExplain": "The ability for AI to use a computer screen to finish tasks.", "humanExplain": "Computer use is like giving the office intern your mouse. They click “OK” before you finish saying, “Wait.”\n\nIt opens websites and clicks buttons. It helps with office work and software tests.", "humanExplainDisplay": "Computer use is like giving the office intern your ==mouse==.\nThey click ==“OK”== before you finish saying, “Wait.”\n\nIt opens websites and clicks buttons.\nIt helps with office work and software tests.", "relationsNarrative": "Agent\nComputer use lets an Agent plan and then act on the computer.\n\nFunction-calling\nComputer use can click the screen when Function-call has no ready API.\n\nCopilot\nA Copilot usually helps from the side. Computer use can operate for you.\n\nData-privacy\nComputer use can see and control the screen, so Data-privacy matters more.", "relations": { "agent": { "label": "lets … act for real", "note": "It turns an Agent from a planner into a clicker and doer." }, "function-call": { "label": "fills gaps in …", "note": "When no API exists, it can click the screen instead." }, "copilot": { "label": "goes beyond …", "note": "A Copilot suggests. Computer use can operate for you." }, "data-privacy": { "label": "raises … risks", "note": "It can see and touch the screen, so privacy risk rises." } } }, "zh": { "fullName": "计算机操作（Computer Use）", "factExplain": "让 AI 直接操作电脑界面完成任务的能力。", "humanExplain": "以前它只会动嘴，现在连鼠标都想接管，活像刚入职就敢替你点确认的实习生。\n\n可直接开网页点按钮，常用于办公自动化、网页操作和测试。", "humanExplainDisplay": "以前它只会动嘴，\n现在连鼠标都想接管，\n活像刚入职就敢替你\n==点确认==的==实习生==。\n\n可直接开网页点按钮，\n常用于办公自动化、\n网页操作和测试。", "relationsNarrative": "Agent\ncomputer use 让 Agent 不只会规划，还能直接在电脑上执行操作。\n\nFunction-calling\n当软件没有现成接口时，computer use 可以直接操作界面来补位。\n\nCopilot\nCopilot 更像在旁建议，computer use 则更进一步，能直接代你操作。\n\nData-privacy\n因为它能看到并操作屏幕内容，所以更容易涉及隐私与权限问题。", "relations": { "agent": { "label": "让…真的动手", "note": "它把 Agent 从会想变成会点会做。" }, "function-call": { "label": "补足…的界面操作", "note": "当没有现成接口时，它可直接点界面。" }, "copilot": { "label": "是…的进阶形态", "note": "Copilot 多是辅助，computer use 更像代操作。" }, "data-privacy": { "label": "更易碰到…问题", "note": "能看见并操作屏幕，隐私风险也更高。" } } } } }, { "id": "computer-vision", "name": "Computer Vision", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "alexnet" }, { "to": "resnet" }, { "to": "multimodal" } ], "track": "history", "i18n": { "en": { "fullName": "Computer Vision", "factExplain": "A field that helps machines understand images and videos.", "humanExplain": "Computer vision is like giving a robot eyes and a clue. It can tell a muffin from a chihuahua.\n\nIt powers photo search and security cameras. Self-driving cars use it to read the road.", "humanExplainDisplay": "Computer vision is like giving a robot\n==eyes and a clue==.\nIt can tell a ==muffin from a chihuahua==.\n\nIt powers photo search and security cameras.\nSelf-driving cars use it to read the road.", "relationsNarrative": "CNN\nCNN was once a classic workhorse for vision tasks.\n\nAlexNet\nAlexNet's image contest breakthrough helped modern computer vision take off.\n\nResNet\nResNet made deep networks easier to train, so vision models got deeper.\n\nMultimodal AI\nComputer vision often joins text and speech as part of Multimodal AI.", "relations": { "cnn": { "label": "often uses …", "note": "Early vision tasks used CNNs a lot." }, "alexnet": { "label": "took off with …", "note": "AlexNet sparked the deep learning boom in vision." }, "resnet": { "label": "moved forward with …", "note": "ResNet made very deep vision models easier to train." }, "multimodal": { "label": "expands into …", "note": "Vision now often works with text and speech." } } }, "zh": { "fullName": "计算机视觉", "factExplain": "让机器从图像或视频中理解信息的领域。", "humanExplain": "计算机视觉不是给机器装摄像头就完事，而是让它像交警一样一眼看懂车、人和路况。\n\n它用于识图、监控分析和自动驾驶，是机器理解画面的基础能力。", "humanExplainDisplay": "计算机视觉\n不是给机器装摄像头就完事，\n而是让它像==交警==一样\n==一眼看懂==车、人和路况。\n\n它用于识图、\n监控分析和自动驾驶，\n是机器理解画面的基础能力。", "relationsNarrative": "CNN\nCNN 曾是视觉任务里最经典的主力模型之一。\n\nAlexNet\nAlexNet 在图像竞赛上的突破带火了现代计算机视觉。\n\nResNet\nResNet 解决深层网络训练难题，推动视觉模型继续变深。\n\nMultimodal AI\n计算机视觉常与文本、语音结合，成为多模态能力的一部分。", "relations": { "cnn": { "label": "常用…建模", "note": "早期视觉任务大量依赖 CNN。" }, "alexnet": { "label": "被…带火", "note": "AlexNet 让深度视觉大爆发。" }, "resnet": { "label": "被…推进", "note": "ResNet 让更深视觉模型可训练。" }, "multimodal": { "label": "扩展到…", "note": "视觉如今常和文字语音一起用。" } } } } }, { "id": "conditional-random-field", "name": "CRF", "layer": "L2", "era": "2001", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "probabilistic-graphical-model" }, { "to": "sequence-labeling" }, { "to": "hidden-markov-model" }, { "to": "named-entity-recognition" } ], "track": "history", "i18n": { "en": { "fullName": "Conditional Random Field", "factExplain": "A model that predicts labels by checking nearby labels too.", "humanExplain": "CRF is like labeling kids in a school lunch line. It checks each kid, then checks if nearby name tags make sense.\n\nIt helps AI label each word in a sentence. You meet it in word splitting, grammar tags, and name finding.", "humanExplainDisplay": "CRF is like labeling kids\nin a ==school lunch line==.\nIt checks each kid,\nthen checks if nearby name tags ==make sense==.\n\nIt helps AI label each word\nin a sentence.\nYou meet it in word splitting,\ngrammar tags,\nand name finding.", "relationsNarrative": "PGM\nCRF is a discriminative PGM that predicts label probabilities directly.\n\nSequence Labeling\nCRF is often used for Sequence Labeling, so nearby labels fit together.\n\nHMM\nCRF makes fewer independence assumptions than HMM and predicts labels directly.\n\nNER\nNER uses CRF to keep entity boundaries clear and smooth.", "relations": { "probabilistic-graphical-model": { "label": "is a kind of …", "note": "A CRF is a discriminative PGM for labels." }, "sequence-labeling": { "label": "helps with …", "note": "CRF labels items in a sequence one by one." }, "hidden-markov-model": { "label": "improves on …", "note": "HMM models observations. CRF predicts labels directly." }, "named-entity-recognition": { "label": "used for …", "note": "CRF helps NER keep entity boundaries smooth." } } }, "zh": { "fullName": "条件随机场", "factExplain": "一种建模相邻标签依赖的判别式概率模型。", "humanExplain": "CRF像排队报数贴胸牌：不只看本人，也看前后号能不能连上。\n\n做分词、词性、实体识别，让标签序列更顺。", "humanExplainDisplay": "CRF像排队报数贴胸牌：\n不只看本人，\n也看前后号\n==能不能连上==。\n\n做分词、词性、实体识别，\n让标签序列\n更顺。", "relationsNarrative": "PGM\nCRF 是判别式概率图模型，直接建模标签条件概率。\n\nSequence Labeling\nCRF 常用于序列标注，让相邻标签彼此配合。\n\nHMM\n它比 HMM 更少假设观测独立，常直接判标签。\n\nNER\nNER 用它约束实体边界，让结果更连贯。", "relations": { "probabilistic-graphical-model": { "label": "属于…", "note": "CRF 是判别式概率图模型。" }, "sequence-labeling": { "label": "服务于…", "note": "它常用来给序列逐项贴标签。" }, "hidden-markov-model": { "label": "改进…", "note": "HMM 生成观测，CRF 直接判标签。" }, "named-entity-recognition": { "label": "用于…", "note": "NER 常用它保证实体边界连贯。" } } } } }, { "id": "conflict-driven-clause-learning", "name": "CDCL", "layer": "L2", "era": "1996", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "dpll-algorithm" }, { "to": "constraint-satisfaction-problem" }, { "to": "resolution-principle" }, { "to": "automated-theorem-proving" } ], "track": "history", "i18n": { "en": { "fullName": "Conflict-Driven Clause Learning", "factExplain": "A SAT solver method that learns extra constraints from conflicts.", "humanExplain": "CDCL is like a kid in a corn maze with a marker. Every dead end gets a big NOPE sign, so whole paths get skipped.\n\nIt powers SAT solvers. It helps check designs and plan steps by turning dead ends into shortcuts.", "humanExplainDisplay": "CDCL is like a kid in a ==corn maze== with a marker.\nEvery dead end gets a big ==NOPE sign==,\nso whole paths get skipped.\n\nIt powers SAT solvers.\nIt helps check designs and plan steps\nby turning dead ends into shortcuts.", "relationsNarrative": "DPLL\nCDCL adds conflict learning to DPLL backtracking search.\n\nCSP\nSAT is a Boolean CSP, and CDCL is built to solve it.\n\nResolution\nA new clause learned by CDCL can act like a Resolution step.\n\nATP\nModern ATP often hands SAT solving to CDCL.", "relations": { "dpll-algorithm": { "label": "extends …", "note": "It adds conflict learning to DPLL search." }, "constraint-satisfaction-problem": { "label": "solves …", "note": "SAT is a Boolean kind of CSP." }, "resolution-principle": { "label": "borrows from …", "note": "A learned clause can be seen as a Resolution step." }, "automated-theorem-proving": { "label": "supports …", "note": "ATP tools often call SAT solvers inside." } } }, "zh": { "fullName": "Conflict-Driven Clause Learning，冲突驱动子句学习", "factExplain": "一种从冲突中学习约束的 SAT 求解算法。", "humanExplain": "CDCL 像扫雷老手贴红旗：踩到雷就记规律，后面整片格子直接避开。\n\n用于 SAT、验证和规划，把死胡同变成剪枝线索。", "humanExplainDisplay": "CDCL 像==扫雷老手贴红旗==：\n==踩到雷就记规律==，\n后面整片格子\n直接避开。\n\n用于 SAT、验证和规划，\n把死胡同，\n变成剪枝线索。", "relationsNarrative": "DPLL\nCDCL 在 DPLL 回溯搜索上加入冲突学习。\n\nCSP\nSAT 可看作布尔约束满足问题，CDCL 专门求解它。\n\nResolution\nCDCL 学到的新子句，可视作一次归结推理。\n\nATP\n现代自动定理证明常把 SAT 求解交给 CDCL。", "relations": { "dpll-algorithm": { "label": "扩展…", "note": "在 DPLL 搜索中加入冲突学习。" }, "constraint-satisfaction-problem": { "label": "求解…", "note": "SAT 是布尔约束满足问题。" }, "resolution-principle": { "label": "借用…", "note": "学到的子句可视作归结结果。" }, "automated-theorem-proving": { "label": "支撑…", "note": "定理证明常调用 SAT 求解器。" } } } } }, { "id": "connectionism", "name": "Connectionism", "layer": "L1", "era": "1980", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "deep-learning" }, { "to": "backpropagation" } ], "track": "history", "i18n": { "en": { "fullName": "Connectionism", "factExplain": "A view that intelligence can grow from learned links between neuron-like units.", "humanExplain": "Connectionism says intelligence works like a school cafeteria at lunch. Kids switch seats, and friend groups pop up by themselves.\n\nIt is the big idea behind neural networks. It later shaped deep learning and large language models.", "humanExplainDisplay": "Connectionism says intelligence works like\na ==school cafeteria at lunch==.\nKids switch seats,\nand ==friend groups pop up== by themselves.\n\nIt is the big idea behind neural networks.\nIt later shaped deep learning\nand large language models.", "relationsNarrative": "Neural-network\nConnectionism gave Neural-network its core idea.\n\nDeep Learning\nDeep Learning is connectionism after computers got much stronger.\n\nBackpropagation\nBackprop lets connectionist systems learn useful links.", "relations": { "neural-network": { "label": "inspires …", "note": "It sees intelligence as many links working together." }, "deep-learning": { "label": "sets up …", "note": "Deep learning is the modern version of this path." }, "backpropagation": { "label": "trains through …", "note": "Backprop lets the links learn from errors." } } }, "zh": { "fullName": "联结主义", "factExplain": "一种认为智能可由神经元连接学习形成的范式。", "humanExplain": "它不信脑子里藏着本天书，更像夜市人流：摊主、顾客来回串，热闹着热闹着秩序自己出来了。\n\n它是神经网络的思想底座，后来一路影响到深度学习和大模型。", "humanExplainDisplay": "它不信脑子里藏着本==天书==，\n更像夜市人流：摊主、顾客来回串，\n热闹着热闹着\n==秩序自己出来了==。\n\n它是神经网络的思想底座，\n后来一路影响到深度学习和大模型。", "relationsNarrative": "Neural-network\n它的核心主张，后来被神经网络具体化了。\n\nDeep Learning\n深度学习是联结主义在大算力时代的爆发。\n\nBackpropagation\n误差反向传播让这套想法终于能有效训练。", "relations": { "neural-network": { "label": "启发…形式", "note": "它把智能看成大量连接共同作用。" }, "deep-learning": { "label": "奠定…思想", "note": "深度学习是这一路线的现代做法。" }, "backpropagation": { "label": "依靠…训练", "note": "误差反传让连接能被真正学出来。" } } } } }, { "id": "connectionist-temporal-classification", "name": "CTC", "layer": "L2", "era": "2006", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "sequence-modeling" }, { "to": "lstm" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Connectionist Temporal Classification", "factExplain": "A training goal for labels when input and output timing does not match.", "humanExplain": "CTC is like a drive-thru worker hearing a muffled burger order. It does not need each word on a stopwatch to get the order right.\n\nIt is common in speech-to-text. It lets models train when sounds and words do not line up neatly.", "humanExplainDisplay": "CTC is like a ==drive-thru worker==\nhearing a muffled burger order.\nIt does not need each word on a ==stopwatch==\nto get the order right.\n\nIt is common in speech-to-text.\nIt lets models train\nwhen sounds and words do not line up neatly.", "relationsNarrative": "STT\nCTC is one classic way to train speech-to-text.\n\nSequence Modeling\nCTC handles tasks where input and output do not line up neatly.\n\nLSTM\nBefore Transformers became common, CTC often worked with LSTM speech models.\n\nClassification\nCTC is still classification, but the labels run across time.", "relations": { "speech-to-text": { "label": "trains …", "note": "CTC is a classic training goal for speech recognition." }, "sequence-modeling": { "label": "handles alignment in …", "note": "CTC helps when input steps and output labels do not match." }, "lstm": { "label": "often paired with …", "note": "Older speech systems often used CTC with LSTM models." }, "classification": { "label": "is a sequence form of …", "note": "CTC is still classification, but across a sequence." } } }, "zh": { "fullName": "Connectionist Temporal Classification｜连接时序分类（CTC）", "factExplain": "一种用于未对齐序列标注的训练目标。", "humanExplain": "CTC 像医生听你含糊报症状，不必逐字卡时间，也能把整段话整理成病历。\n\n常用于语音转文字这类任务，输入输出对不上号时，也能让模型正常训练。", "humanExplainDisplay": "CTC 像医生听你含糊报症状，\n不必==逐字卡时间==，\n也能把整段话==整理成病历==。\n\n常用于语音转文字这类任务，\n输入输出对不上号时，\n也能让模型正常训练。", "relationsNarrative": "Speech-to-text\n它是语音识别里最经典的训练方法之一。\n\nSequence Modeling\n它专门处理序列任务里输入输出难对齐的问题。\n\nLSTM\n在 Transformer 普及前，它常和 LSTM 一起用于语音任务。\n\nClassification\n它本质上仍是分类，只是分类对象变成了时序序列。", "relations": { "speech-to-text": { "label": "常用于…训练", "note": "它是语音识别里的经典训练目标。" }, "sequence-modeling": { "label": "处理…对齐难题", "note": "专门解决序列输入输出难对齐。" }, "lstm": { "label": "常搭配…使用", "note": "早期常和 LSTM 语音模型配套。" }, "classification": { "label": "属于…变体", "note": "本质仍是在做时序上的分类。" } } } } }, { "id": "constitutional-ai", "name": "Constitutional AI", "layer": "L6", "era": "2022", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "alignment" }, { "to": "rlhf" }, { "to": "post-training" } ], "track": "history", "i18n": { "en": { "fullName": "Constitutional AI", "factExplain": "A training method where AI uses rules to critique and fix its own answers.", "humanExplain": "Constitutional AI is like a family rule sheet on the fridge. Before the AI blurts out a wild answer, it checks the rules and fixes itself.\n\nYou meet it in safety training for chatbots. It can reduce harmful replies and need fewer human graders.", "humanExplainDisplay": "Constitutional AI is like a ==family rule sheet==\non the fridge.\nBefore the AI blurts out a ==wild answer==,\nit checks the rules\nand fixes itself.\n\nYou meet it in safety training for chatbots.\nIt can reduce harmful replies\nand need fewer human graders.", "relationsNarrative": "Alignment\nConstitutional AI writes values into the model's behavior.\n\nRLHF\nConstitutional AI uses rules and AI feedback, so it needs less human feedback.\n\nPost-training\nConstitutional AI usually happens after pre-training, during model tuning.", "relations": { "alignment": { "label": "serves …", "note": "It turns safety principles into limits on model behavior." }, "rlhf": { "label": "relies less on …", "note": "Principles and AI feedback replace some human feedback." }, "post-training": { "label": "is part of …", "note": "It is usually safety tuning after pre-training." } } }, "zh": { "fullName": "宪法式人工智能", "factExplain": "用原则让模型自我批评并修正回答的训练方法。", "humanExplain": "宪法式 AI 像小区公约：开口先看规矩，发现越界就自查改口。\n\n用于大模型安全对齐，减少有害回答和人工标注。", "humanExplainDisplay": "宪法式 AI 像==小区公约==：\n开口先看规矩，\n发现越界，\n就==自查改口==。\n\n用于大模型安全对齐，\n减少有害回答\n和人工标注。", "relationsNarrative": "Alignment\n它是把价值原则写进模型行为的对齐方法。\n\nRLHF\n它用原则和 AI 反馈，减少对人工反馈的依赖。\n\nPost-training\n它通常发生在预训练之后，属于模型调校阶段。", "relations": { "alignment": { "label": "服务…", "note": "把原则变成模型行为约束。" }, "rlhf": { "label": "减少对…依赖", "note": "用原则和 AI 反馈替代部分人工反馈。" }, "post-training": { "label": "属于…", "note": "多在预训练后做安全调校。" } } } } }, { "id": "constraint-satisfaction-problem", "name": "CSP", "layer": "L2", "era": "1970s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "optimization" }, { "to": "heuristic-search" }, { "to": "graph-search" }, { "to": "knowledge-representation" } ], "track": "history", "i18n": { "en": { "fullName": "Constraint Satisfaction Problem", "factExplain": "A problem where you find any answer that follows all the rules.", "humanExplain": "CSP is a wedding seating chart with drama. Bob avoids Sue. Grandma gets the aisle. Dinner survives.\n\nIt shows up in timetables and work schedules. The goal is a valid answer first, not the best one.", "humanExplainDisplay": "CSP is a ==wedding seating chart== with ==drama==.\nBob avoids Sue.\nGrandma gets the aisle.\nDinner survives.\n\nIt shows up in timetables and work schedules.\nThe goal is a valid answer first,\nnot the best one.", "relationsNarrative": "Optimization\nCSP first looks for an answer that follows all rules. Optimization looks for a better answer.\n\nHeuristic Search\nMany CSPs use heuristic search to find a valid answer faster.\n\nGraph Search\nA CSP can become graph search when partial answers are treated as states.\n\nKR\nKR states the rules clearly, so the system knows which answers are allowed.", "relations": { "optimization": { "label": "sits beside …", "note": "CSP finds a valid answer. Optimization keeps looking for the best one." }, "heuristic-search": { "label": "often solved with …", "note": "Heuristics help a solver shrink the search faster." }, "graph-search": { "label": "can become …", "note": "Treat each partial answer as a state, then search through states." }, "knowledge-representation": { "label": "needs … to state rules", "note": "The system needs clear variables and rules before it can judge answers." } } }, "zh": { "fullName": "Constraint Satisfaction Problem（约束满足问题）", "factExplain": "在给定约束下寻找可行解的问题。", "humanExplain": "像排喜宴座位表：这俩有过节不能同桌、长辈得坐主桌、伴娘要挨着新娘，所有规矩都迁就到，人人落座不闹掰就算成。\n\n常见于安排和调度，重点是先找到满足条件的解。", "humanExplainDisplay": "像排喜宴==座位表==：\n这俩有过节不能同桌、\n长辈得坐主桌、\n伴娘要挨着新娘，\n所有规矩都迁就到，\n人人落座==不闹掰==就算成。\n\n常见于安排和调度，\n重点是先找到\n满足条件的解。", "relationsNarrative": "Optimization\n约束满足先求满足条件，优化问题还会继续追求更优。\n\nHeuristic Search\n很多约束问题会靠启发式搜索更快找到可行解。\n\nGraph Search\n把部分解看成状态节点，就能转成搜索问题。\n\nKnowledge Representation\n约束得先被清楚表示，系统才知道哪些解合法。", "relations": { "optimization": { "label": "常与…并列", "note": "前者找可行解，后者常追求最优。" }, "heuristic-search": { "label": "常靠…求解", "note": "很多求解器靠启发式更快缩小搜索。" }, "graph-search": { "label": "可转成…", "note": "把状态当节点，就能按搜索方式求解。" }, "knowledge-representation": { "label": "用…表达约束", "note": "变量、规则和关系都需先形式化。" } } } } }, { "id": "content-provenance", "name": "Content provenance", "layer": "L6", "era": "2023", "publishedAt": "2026-05-28T15:58:23.414Z", "relations": [ { "to": "deepfake" }, { "to": "copyright" }, { "to": "ai-regulation" }, { "to": "ai-anxiety" } ], "track": "today", "i18n": { "en": { "fullName": "Content provenance", "factExplain": "A way to mark and check where content came from and how it changed.", "humanExplain": "Content provenance is a luggage tag for a video. It says who packed it and what stops it made.\n\nYou meet it in news checks and social apps. It helps spot AI-made content and judge trust.", "humanExplainDisplay": "Content provenance is a ==luggage tag== for a video.\nIt says who packed it\nand ==what stops it made==.\n\nYou meet it in news checks\nand social apps.\nIt helps spot AI-made content\nand judge trust.", "relationsNarrative": "Deepfake\nContent provenance gives clues about a Deepfake, but it cannot prove truth alone.\n\nCopyright\nContent provenance records how a work was made and changed, so ownership is easier to judge.\n\nAI-regulation\nAI-regulation often asks platforms to track where content came from.\n\nAI-anxiety\nAI-anxiety often drops when people can see where content came from.", "relations": { "deepfake": { "label": "helps spot …", "note": "It gives source clues when people argue if media is fake." }, "copyright": { "label": "helps judge … ownership", "note": "It adds records of how a work was made and changed." }, "ai-regulation": { "label": "is often required by …", "note": "Platform rules often ask content sources to be traceable." }, "ai-anxiety": { "label": "eases … worries", "note": "People feel safer when they can see where content came from." } } }, "zh": { "fullName": "内容来源追踪", "factExplain": "用于标记和验证内容来源与编辑历史的机制。", "humanExplain": "给内容补一张来路清单，谁生的、谁改的、怎么传的，都别再靠大家集体猜谜。\n\n用于新闻核验、平台溯源和识别 AI 内容，帮助判断可信度。", "humanExplainDisplay": "给内容补一张\n==来路清单==，\n谁生的、谁改的、怎么传的，\n都别再靠大家\n==集体猜谜==。\n\n用于新闻核验、\n平台溯源和识别 AI 内容，\n帮助判断可信度。", "relationsNarrative": "Deepfake\nContent provenance 能为 Deepfake 的真伪争议提供来源线索，但不能单独保证内容一定真实。\n\nCopyright\nContent provenance 可补充内容的生成与修改记录，帮助判断作品归属与使用边界。\n\nAI-regulation\nAI-regulation 常把内容来源可追踪性当作平台治理和披露要求的一部分。\n\nAI-anxiety\n当人们能看到内容从哪来、怎么来的，AI-anxiety 往往会降低一些。", "relations": { "deepfake": { "label": "帮助识别…", "note": "可为真假争议提供来源线索。" }, "copyright": { "label": "辅助判断…归属", "note": "能补充作品生成与修改记录。" }, "ai-regulation": { "label": "常被纳入…要求", "note": "平台治理常要求可追踪来源。" }, "ai-anxiety": { "label": "缓解…不安", "note": "知道内容来路，用户更敢判断。" } } } } }, { "id": "context-compression", "name": "Ctx Compression", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "context-window" }, { "to": "rag" }, { "to": "agent-memory" }, { "to": "token" } ], "track": "today", "i18n": { "en": { "fullName": "Context Compression", "factExplain": "A way to shorten input text while keeping the key information.", "humanExplain": "Context Compression is like a friend recapping a three-hour movie in the popcorn line. You get the dragon fight, not the carpet color.\n\nIt shrinks long text before the AI reads it. You meet it in long document Q&A. It also helps Memory and huge chats.", "humanExplainDisplay": "Context Compression is like a friend recapping\n==a three-hour movie== in the popcorn line.\nYou get the ==dragon fight==,\nnot the carpet color.\n\nIt shrinks long text before the AI reads it.\nYou meet it in long document Q&A.\nIt also helps Memory and huge chats.", "relationsNarrative": "Context-window\nContext Compression uses shorter text to fit more information into the window.\n\nRAG\nRAG often uses it to shorten found sources before giving them to the model.\n\nMemory\nIt turns long past memory into shorter information the AI can still use.\n\nToken\nContext Compression reduces the number of input tokens.", "relations": { "context-window": { "label": "fits more into …", "note": "It shrinks text before it enters the limited window." }, "rag": { "label": "works with …", "note": "RAG finds sources, then compression trims them for the model." }, "agent-memory": { "label": "tidies …", "note": "It turns long memory into a short, useful summary." }, "token": { "label": "uses fewer …", "note": "A main goal is to spend fewer input tokens." } } }, "zh": { "fullName": "Context Compression（上下文压缩）", "factExplain": "在尽量保留关键信息下缩短输入上下文的方法。", "humanExplain": "像朋友排队时一句话复述三小时电影：==打斗高潮==留给你，==地毯啥颜色==这种细节直接略过。\n\n常用于长文档问答、Agent 记忆和超长对话，帮模型省上下文。", "humanExplainDisplay": "像朋友排队时一句话复述三小时电影：\n==打斗高潮==留给你，\n==地毯啥颜色==这种细节直接略过。\n\n常用于长文档问答、\nAgent 记忆和超长对话，\n帮模型省上下文。", "relationsNarrative": "Context-window\n它用更短文本装下更多信息，缓解窗口限制。\n\nRAG\nRAG 找资料后，常靠它把重点压短再喂给模型。\n\nMemory\n它能把冗长历史记忆整理成更紧凑的可用信息。\n\nToken\n压缩上下文，本质上是在减少输入 token 消耗。", "relations": { "context-window": { "label": "缓解…不够用", "note": "先压缩内容，再塞进有限窗口。" }, "rag": { "label": "常配合…使用", "note": "先找资料，再压缩重点给模型。" }, "agent-memory": { "label": "整理…内容", "note": "把长记忆浓缩成可继续用的摘要。" }, "token": { "label": "减少…占用", "note": "核心目标之一是少花输入 token。" } } } } }, { "id": "context-window", "name": "Context-window", "layer": "L2", "era": "2023", "publishedAt": "2026-05-23T08:25:00Z", "relations": [ { "to": "token" }, { "to": "llm" }, { "to": "prompt" } ], "i18n": { "en": { "fullName": "Context Window", "factExplain": "The amount of chat text an AI can read at one time.", "humanExplain": "A context window is the AI’s kitchen counter. Bigger counter, more notes and pizza boxes fit.\n\nIt sets how much earlier chat the AI can use. When it fills up, older text may get pushed out.", "humanExplainDisplay": "A context window is the AI’s ==kitchen counter==.\nBigger counter,\nmore ==notes and pizza boxes== fit.\n\nIt sets how much earlier chat\nthe AI can use.\nWhen it fills up,\nolder text may get pushed out.", "relationsNarrative": "Token\nA context window size is usually measured by how many tokens fit inside.\n\nLLM\nAn LLM can only answer from information inside its context window.\n\nPrompt\nA longer prompt takes up more space in the context window.", "relations": { "token": { "label": "is measured in …", "note": "A context window size is usually counted in tokens." }, "llm": { "label": "sets memory for …", "note": "An LLM can only use text inside its context window." }, "prompt": { "label": "limits … length", "note": "A longer prompt uses more room in the context window." } } }, "zh": { "fullName": "上下文窗口", "factExplain": "模型一次对话中能够读取和参考的信息范围。", "humanExplain": "上下文窗口像 AI 的临时办公桌，桌面越大，能摊开的资料越多。\n\n它决定 AI 能记住多少前文；桌子满了，旧东西就可能被挤下去。", "humanExplainDisplay": "上下文窗口像 AI 的\n==临时办公桌==。\n桌面越大，能摊开的资料越多。\n\n前文、文件、提示词都会占位置。\n桌子满了，旧信息就可能被挤下去。", "relationsNarrative": "Token\nContext-window 的容量通常用可容纳 Token 数衡量。\n\nLLM\nLLM 只能基于 Context-window 内的信息生成回答。\n\nPrompt\n更长的 Prompt 会占用更多 Context-window 空间。", "relations": { "token": { "label": "由…数限制" }, "llm": { "label": "决定…的记忆" }, "prompt": { "label": "限制…长度" } } } } }, { "id": "continual-learning", "name": "Continual Learning", "layer": "L2", "era": "1990s", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "online-learning" }, { "to": "transfer-learning" }, { "to": "fine-tuning" }, { "to": "multitask-learning" } ], "track": "today", "i18n": { "en": { "fullName": "Continual Learning", "factExplain": "A way for AI to learn new tasks without forgetting old ones.", "humanExplain": "Continual Learning is like a cafeteria cook learning taco Tuesday. Pizza Friday must still taste like pizza.\n\nIt helps long-term assistants and robots keep improving. Recommendation apps use it too, without wiping old taste clues.", "humanExplainDisplay": "Continual Learning is like a cafeteria cook\nlearning ==taco Tuesday==.\n==Pizza Friday== must still taste like pizza.\n\nIt helps long-term assistants and robots\nkeep improving.\nRecommendation apps use it too,\nwithout wiping old taste clues.", "relationsNarrative": "Online Learning\nContinual Learning often updates the model as new data arrives.\n\nTransfer Learning\nTransfer Learning brings old task experience into new tasks.\n\nFine-tuning\nFine-tuning can teach new data, but it may wipe old skills.\n\nMultitask Learning\nMultitask Learning learns tasks together. Continual Learning learns them in order.", "relations": { "online-learning": { "label": "often works with …", "note": "Online Learning updates the model as new data arrives." }, "transfer-learning": { "label": "builds on …", "note": "Transfer Learning carries old experience into a new task." }, "fine-tuning": { "label": "can use …", "note": "Fine-tuning can teach new data, but may overwrite old skills." }, "multitask-learning": { "label": "differs from …", "note": "Multitask Learning trains several tasks at the same time." } } }, "zh": { "fullName": "持续学习", "factExplain": "让模型连续学习新任务且尽量不遗忘旧知识。", "humanExplain": "持续学习像武馆师傅加练新招：会了咏春，也不能把太极还给师父。\n\n用于长期助手、机器人、推荐系统，边更新边少遗忘。", "humanExplainDisplay": "持续学习像武馆师傅\n==加练新招==：\n会了咏春，\n也不能==把太极还给师父==。\n\n用于长期助手、\n机器人、推荐系统，\n边更新边少遗忘。", "relationsNarrative": "Online Learning\n在线学习强调数据不断到来时持续更新模型。\n\nTransfer Learning\n迁移学习把旧任务经验迁到新任务，持续学习更重防遗忘。\n\nFine-tuning\n微调可让模型学新数据，但也可能冲掉旧能力。\n\nMultitask Learning\n多任务学习多是一起学，持续学习强调按顺序学。", "relations": { "online-learning": { "label": "常结合…", "note": "在线学习强调边来数据边更新。" }, "transfer-learning": { "label": "延续…的思路", "note": "迁移学习把旧经验带到新任务。" }, "fine-tuning": { "label": "可通过…实现", "note": "微调常用于适配新数据或新任务。" }, "multitask-learning": { "label": "区别于…", "note": "多任务学习通常同时训练多个任务。" } } } } }, { "id": "continuous-batching", "name": "Continuous batching", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-05-28T15:58:23.420Z", "relations": [ { "to": "inference" }, { "to": "gpu" }, { "to": "token" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Continuous batching", "factExplain": "A scheduling method for adding new requests to a batch already running.", "humanExplain": "Continuous batching is a diner grill at lunch rush. New burgers slide on while the cook flips the first batch.\n\nYou meet it in online LLM inference. It keeps the GPU busy. Answers can arrive faster.", "humanExplainDisplay": "Continuous batching is a ==diner grill==\nat lunch rush.\n==New burgers slide on==\nwhile the cook flips the first batch.\n\nYou meet it in online LLM inference.\nIt keeps the GPU busy.\nAnswers can arrive faster.", "relationsNarrative": "Inference\nContinuous batching is used during inference to run requests together more efficiently.\n\nGPU\nContinuous batching cuts GPU waiting time and keeps compute packed tighter.\n\nToken\nLLMs produce tokens step by step, so new requests can join during generation.\n\nLLM\nContinuous batching is common in live LLM services, especially with many users at once.", "relations": { "inference": { "label": "optimizes … scheduling", "note": "It mainly happens while the model is making answers." }, "gpu": { "label": "cuts … idle time", "note": "It keeps the GPU working instead of waiting around." }, "token": { "label": "batches by … steps", "note": "Tokens come out step by step, so new requests can join between steps." }, "llm": { "label": "often serves …", "note": "It is common in live LLM chat services." } } }, "zh": { "fullName": "连续批处理（Continuous Batching）", "factExplain": "让请求在运行中持续并入同批处理的调度方式。", "humanExplain": "它像早高峰拼车：不等坐满才发，新乘客随时上，空座马上补。\n\n它常用于大模型在线推理，让 GPU 少空等、服务更扛流量。", "humanExplainDisplay": "它像==早高峰拼车==：\n不等坐满才发，\n新乘客随时上，\n==空座马上补==。\n\n它常用于大模型在线推理，\n让 GPU 少空等、\n服务更扛流量。", "relationsNarrative": "Inference\ncontinuous batching 主要用于推理阶段，解决请求怎么一起跑更划算。\n\nGPU\n它能减少 GPU 等活的时间，让算力利用更紧凑。\n\nToken\n因为回答是一段段往外吐，所以新请求能在生成过程中并入。\n\nLLM\n它最常见于 LLM 在线服务，尤其是多人同时请求时。", "relations": { "inference": { "label": "优化…调度", "note": "它主要发生在模型实际出结果时。" }, "gpu": { "label": "减少…空等", "note": "让算力利用更紧凑，不必频繁闲置。" }, "token": { "label": "按…续批", "note": "生成是逐步吐出的，适合动态插队。" }, "llm": { "label": "常用于…服务", "note": "大模型在线问答里最常见这套办法。" } } } } }, { "id": "contrastive-learning", "name": "Contrastive Learning", "layer": "L2", "era": "2006", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "self-supervised-learning" }, { "to": "embedding" }, { "to": "clip" }, { "to": "multimodal" } ], "track": "history", "i18n": { "en": { "fullName": "Contrastive Learning", "factExplain": "A way to learn by pulling similar examples close and pushing different ones apart.", "humanExplain": "Contrastive learning is a school cafeteria seating chart. Best friends share fries. Frenemies get the far table.\n\nIt teaches AI useful Embeddings for search and matching. It also helps pictures and captions meet in the same space.", "humanExplainDisplay": "Contrastive learning is a ==school cafeteria seating chart==.\n==Best friends share fries==.\nFrenemies get the far table.\n\nIt teaches AI useful Embeddings\nfor search and matching.\nIt also helps pictures and captions\nmeet in the same space.", "relationsNarrative": "SSL\nContrastive learning is a common SSL method.\n\nEmbedding\nIt often makes better Embeddings.\n\nCLIP\nCLIP uses it to map images and text into one space.\n\nMultimodal AI\nIt helps match pictures, text, and other data types.", "relations": { "self-supervised-learning": { "label": "is a common method in …", "note": "It can learn similarity from unlabeled data." }, "embedding": { "label": "trains …", "note": "It pulls similar vectors together and pushes others apart." }, "clip": { "label": "helps … align", "note": "CLIP uses it to pull matching images and text together." }, "multimodal": { "label": "supports … alignment", "note": "It helps different data types find matching pairs." } } }, "zh": { "fullName": "对比学习（Contrastive Learning）", "factExplain": "一种通过拉近相似样本、拉远不同样本来学表示的方法。", "humanExplain": "它像相亲局排座位：看对眼的往一桌挪，聊不来的别硬撮，排久了，谁跟谁合拍就越来越清楚。\n\n它常用来学习表示、做检索匹配，也常用于图文等多模态对齐。", "humanExplainDisplay": "它像相亲局排座位：\n看对眼的往==一桌挪==，\n聊不来的别硬撮，\n排久了，谁跟谁==合拍==\n就越来越清楚。\n\n它常用来学习表示、\n做检索匹配，\n也常用于图文等多模态对齐。", "relationsNarrative": "Self-supervised-learning\n对比学习是自监督学习里的代表性方法之一。\n\nEmbedding\n它的核心产物常是更好用的向量表示。\n\nCLIP\nCLIP 用对比学习把图片和文字映射到同一空间。\n\nMultimodal AI\n它常被用来学习图像、文本等模态之间的对应关系。", "relations": { "self-supervised-learning": { "label": "属于…常见方法", "note": "它常靠未标注数据自己学相似性。" }, "embedding": { "label": "训练…表示", "note": "对比目标常直接塑造向量空间。" }, "clip": { "label": "被…用来对齐", "note": "CLIP 靠它把图像和文本拉到一起。" }, "multimodal": { "label": "支撑…对齐", "note": "它常用于不同模态之间找对应。" } } } } }, { "id": "conversational-interface", "name": "Conversational interface", "layer": "L5", "sublayer": "product", "era": "1960s", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "chatbot" }, { "to": "prompt" }, { "to": "natural-language-understanding" }, { "to": "command-line-interface" } ], "track": "today", "i18n": { "en": { "fullName": "Conversational Interface", "factExplain": "An interface that lets you control a system by talking in normal language.", "humanExplain": "A conversational interface is like a hotel front desk. You skip the weird menu maze and just ask, “Where’s the pool?”\n\nYou meet it in support chats, AI helpers, and search. You learn fewer buttons, but the system must understand you.", "humanExplainDisplay": "A conversational interface is like a ==hotel front desk==.\nYou skip the weird menu maze\nand ==just ask==,\n“Where’s the pool?”\n\nYou meet it in support chats,\nAI helpers,\nand search.\nYou learn fewer buttons,\nbut the system must understand you.", "relationsNarrative": "Chatbot\nA conversational interface is the usual shell for a chatbot.\n\nPrompt\nWhat the user says in chat can become the model’s prompt.\n\nNLU\nIt needs NLU to understand the user’s goal.\n\nCommand Line Interface\nIt lowers the Command Line Interface hurdle by using normal talk.", "relations": { "chatbot": { "label": "often appears as …", "note": "A chatbot is a common front door for a conversational interface." }, "prompt": { "label": "turns user talk into …", "note": "A user message can become a model instruction." }, "natural-language-understanding": { "label": "uses … to read intent", "note": "NLU helps the system understand what the user wants." }, "command-line-interface": { "label": "lowers the bar from …", "note": "It turns typed commands into normal talking." } } }, "zh": { "fullName": "对话式界面", "factExplain": "用自然语言对话来操作系统的界面。", "humanExplain": "对话界面像小区门卫大爷：不用填复杂表，跟他说找谁，他就帮你指路。\n\n用于客服、助手和搜索，少学按钮流程，但更考验理解力。", "humanExplainDisplay": "对话界面像==小区门卫大爷==：\n不用填复杂表，\n==跟他说找谁==，\n他就帮你指路。\n\n用于客服、助手和搜索，\n少学按钮流程，\n但更考验理解力。", "relationsNarrative": "Chatbot\n对话界面是聊天机器人最常见的交互外壳。\n\nPrompt\n用户在对话里说的话，常会变成模型指令。\n\nNLU\n它需要理解用户意图，才不像鸡同鸭讲。\n\nCommand Line Interface\n它把敲命令的门槛，降成自然说话。", "relations": { "chatbot": { "label": "常见于…", "note": "聊天机器人是典型对话入口。" }, "prompt": { "label": "把输入变成…", "note": "用户每句话都可能是指令。" }, "natural-language-understanding": { "label": "依赖…识别意图", "note": "先听懂人话，才能顺着聊。" }, "command-line-interface": { "label": "弱化…门槛", "note": "把敲命令改成自然说话。" } } } } }, { "id": "convex-optimization", "name": "Convex optimization", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "optimization" }, { "to": "gradient-descent" }, { "to": "logistic-regression" }, { "to": "support-vector-machine" } ], "track": "history", "i18n": { "en": { "fullName": "Convex Optimization", "factExplain": "A way to optimize when the goal and allowed choices are convex.", "humanExplain": "Think of a marble in a cereal bowl. Let go, and it rolls to the real bottom, not a sneaky pothole.\n\nConvex optimization finds the best choice in bowl-shaped problems with clear rules. You meet it in model training and resource planning.", "humanExplainDisplay": "Think of a ==marble in a cereal bowl==.\nLet go,\nand it rolls to the ==real bottom==,\nnot a sneaky pothole.\n\nConvex optimization finds the best choice\nin bowl-shaped problems with clear rules.\nYou meet it in model training\nand resource planning.", "relationsNarrative": "Optimization\nConvex optimization is a type of optimization with an easier path to the best answer.\n\nGradient Descent\nGradient Descent can solve many convex problems well.\n\nLogit\nLogit training is usually written as a convex optimization problem.\n\nSVM\nClassic SVM training uses a convex goal.", "relations": { "optimization": { "label": "is a special case of …", "note": "It is a large, easier class of optimization problems." }, "gradient-descent": { "label": "is often solved by …", "note": "Gradient Descent can solve many convex problems in a steady way." }, "logistic-regression": { "label": "supports … training", "note": "Logit training is often written as a convex optimization problem." }, "support-vector-machine": { "label": "supports … training", "note": "Classic SVM training depends on a convex goal." } } }, "zh": { "fullName": "凸优化", "factExplain": "研究目标与约束均为凸时的最优化问题。", "humanExplain": "它像在山谷里找最低那口井：你只要顺着坡一直往下走，基本不会掉进假洼地兜圈子。\n\n常用于经典模型训练和资源分配等稳定求最优场景。", "humanExplainDisplay": "它像在山谷里\n找最低那口井：\n你只要顺着坡\n一直==往下走==，\n基本不会掉进\n==假洼地==兜圈子。\n\n常用于经典模型训练，\n和资源分配等\n稳定求最优场景。", "relationsNarrative": "Optimization\n它是最优化中的一类，特点是更容易找到全局最优。\n\nGradient Descent\n很多凸问题能用梯度下降高效求解。\n\nLogistic Regression\n逻辑回归训练通常可表述为凸优化问题。\n\nSVM\n支持向量机的经典训练目标就是凸的。", "relations": { "optimization": { "label": "属于…特例", "note": "它是最优化里最好解的一大类。" }, "gradient-descent": { "label": "常用…求解", "note": "很多凸问题可用梯度下降稳定求解。" }, "logistic-regression": { "label": "支撑…训练", "note": "逻辑回归常写成标准凸优化问题。" }, "support-vector-machine": { "label": "支撑…求解", "note": "支持向量机训练常依赖凸优化。" } } } } }, { "id": "copilot", "name": "Copilot", "layer": "L5", "sublayer": "product", "era": "2021", "publishedAt": "2026-05-23T10:20:00Z", "relations": [ { "to": "llm" }, { "to": "cursor" }, { "to": "vibe-coding" }, { "to": "prompt" } ], "i18n": { "en": { "fullName": "AI Coding Assistant", "factExplain": "An AI tool that helps developers write, explain, and change code.", "humanExplain": "Copilot is like a coding buddy with too much coffee. It types fast, but you still check the work.\n\nIt can write boilerplate code. It can explain a function. It can help fix small bugs. You meet it inside code editors.", "humanExplainDisplay": "Copilot is like a ==coding buddy==\nwith ==too much coffee==.\nIt types fast,\nbut you still check the work.\n\nIt can write boilerplate code.\nIt can explain a function.\nIt can help fix small bugs.\nYou meet it inside code editors.", "relationsNarrative": "LLM\nCopilot uses an LLM to understand code context and suggest code.\n\nCursor\nCursor brings Copilot-like help into a fuller code editor.\n\nVibe-coding\nCopilot helps Vibe-coding grow by turning plain requests into code.\n\nPrompt\nA clearer Prompt helps Copilot write code you can actually use.", "relations": { "llm": { "label": "is powered by …", "note": "Copilot uses an LLM to read code context and suggest code." }, "cursor": { "label": "has examples like …", "note": "Cursor is a code editor with strong Copilot-like help." }, "vibe-coding": { "label": "boosts …", "note": "Copilot makes natural-language coding feel more useful." }, "prompt": { "label": "works better with …", "note": "A clear prompt helps Copilot write more useful code." } } }, "zh": { "fullName": "编程助手", "factExplain": "辅助开发者编写、解释和修改代码的 AI 工具。", "humanExplain": "AI 副驾像坐你旁边的老司机，不抢方向盘，关键时提醒你别开进沟里。\n\n它常嵌在办公、编程、设计工具里，帮人起草、补全和改错。", "humanExplainDisplay": "AI 副驾像==坐你旁边的老司机==，\n不抢方向盘，\n关键时提醒你==别开进沟里==。\n\n它常嵌在办公、编程、设计工具里，\n帮人起草、补全和改错。", "relationsNarrative": "LLM\nCopilot 依靠 LLM 理解代码上下文并生成建议。\n\nCursor\nCursor 将 Copilot 类能力扩展到更完整的开发环境。\n\nVibe-coding\nCopilot 推动了用自然语言辅助开发的 Vibe-coding。\n\nPrompt\nPrompt 越清晰，Copilot 越容易生成可用代码。", "relations": { "llm": { "label": "基于…" }, "cursor": { "label": "代表之一是…" }, "vibe-coding": { "label": "放大…的价值" } } } } }, { "id": "copyright", "name": "Copyright", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T11:05:00Z", "relations": [ { "to": "data-privacy" }, { "to": "deepfake" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "AI and Copyright", "factExplain": "Copyright issues around AI training, AI-made content, and how people use it.", "humanExplain": "AI copyright is like a school art project made from everyone’s stickers. At the end, three kids all yell, “That’s mine!”\n\nIt shows up in training data and AI-made work. It gets serious when a company uses the result. The rules are still catching up.", "humanExplainDisplay": "AI copyright is like a ==school art project==\nmade from everyone’s stickers.\nAt the end,\nthree kids all yell,\n==“That’s mine!”==\n\nIt shows up in training data\nand AI-made work.\nIt gets serious when a company\nuses the result.\nThe rules are still catching up.", "relationsNarrative": "Data-privacy\nData-privacy and Copyright both limit how training data can be used.\n\nDeepfake\nDeepfake makes Copyright fights over faces and ownership louder.\n\nAI-regulation\nAI-regulation turns Copyright fights into rules people can enforce.", "relations": { "data-privacy": { "label": "overlaps with …", "note": "Both limit how training data can be used." }, "deepfake": { "label": "gets louder with …", "note": "Deepfakes make fights over faces and ownership harder." }, "ai-regulation": { "label": "is defined by …", "note": "AI regulation turns copyright fights into rules people can follow." } } }, "zh": { "fullName": "AI 与版权", "factExplain": "AI 训练、生成和使用内容时涉及的著作权问题。", "humanExplain": "AI 版权像一锅大家都往里扔过料的汤，最后谁能说这碗汤归自己，吵起来很正常。\n\n训练数据、生成作品和商业使用都会踩线，需要规则慢慢补课。", "humanExplainDisplay": "AI 版权像一锅\n==大家都往里扔过料的汤==。\n最后谁说这碗汤归自己，\n吵起来很正常。\n\n训练数据、生成作品、\n商业使用都会踩线。\n规则还在补课，\n律师已经开始加班。", "relationsNarrative": "Data-privacy\nData-privacy 和 Copyright 都限制训练数据的使用方式。\n\nDeepfake\nDeepfake 放大了 Copyright 中关于肖像和内容归属的争议。\n\nAI-regulation\nAI-regulation 将 Copyright 争议转化为可执行规则。", "relations": { "data-privacy": { "label": "涉及…" }, "deepfake": { "label": "被…放大" }, "ai-regulation": { "label": "由…界定" } } } } }, { "id": "coreference-resolution", "name": "Coreference Resolution", "layer": "L4", "era": "1970", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "natural-language-understanding" }, { "to": "information-extraction" }, { "to": "named-entity-recognition" } ], "track": "history", "i18n": { "en": { "fullName": "Coreference Resolution", "factExplain": "A way to link words in text to the same person or thing.", "humanExplain": "Coreference resolution is the name-tag helper at a family reunion. It spots when “she” and “Aunt Mia” are the same person.\n\nIt keeps Q&A and summaries from mixing people up. It also helps tools pull facts from text.", "humanExplainDisplay": "Coreference resolution is the ==name-tag helper==\nat a family reunion.\nIt spots when ==“she” and “Aunt Mia”==\nare the same person.\n\nIt keeps Q&A and summaries\nfrom mixing people up.\nIt also helps tools\npull facts from text.", "relationsNarrative": "NLU\nCoreference resolution helps NLU know who pronouns and names point to.\n\nIE\nIE uses coreference resolution to merge different mentions of the same object.\n\nNER\nNER finds entities first, then coreference resolution checks which ones match.", "relations": { "natural-language-understanding": { "label": "helps … track references", "note": "Coreference resolution helps NLU know who each word points to." }, "information-extraction": { "label": "helps … merge entities", "note": "IE uses it to join different names for the same thing." }, "named-entity-recognition": { "label": "links after … finds names", "note": "NER finds the names first; coreference resolution connects matches." } } }, "zh": { "fullName": "共指消解", "factExplain": "识别文本中指向同一对象的表达。", "humanExplain": "共指消解像相亲局对暗号：“他”“老张”“孩子爸”，得认出都是同一人。\n\n让问答、摘要和信息抽取少串人，长文理解更稳。", "humanExplainDisplay": "共指消解像\n==相亲局对暗号==：\n“他”“老张”“孩子爸”，\n得认出==都是同一人==。\n\n让问答、摘要和信息抽取\n少串人，\n长文理解更稳。", "relationsNarrative": "NLU\n共指消解让系统理解代词和名称指向谁。\n\nIE\n信息抽取常用它把同一对象的多种说法合并。\n\nNER\nNER 先找出实体，共指消解再判断谁和谁相同。", "relations": { "natural-language-understanding": { "label": "支撑…理解指代", "note": "共指消解让系统读懂谁在指谁。" }, "information-extraction": { "label": "帮…合并实体", "note": "同一对象的多种说法需要归到一起。" }, "named-entity-recognition": { "label": "接在…之后连线", "note": "先认出名字，再判断是否同一人。" } } } } }, { "id": "cost-aware-ai-ai", "name": "Cost-aware AI", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "model-routing" }, { "to": "reasoning-effort" }, { "to": "llmops" }, { "to": "api" } ], "track": "today", "i18n": { "en": { "fullName": "Cost-aware AI", "factExplain": "AI that weighs cost before choosing models, steps, or API calls.", "humanExplain": "Cost-aware AI is the parent packing school lunch. Not every Tuesday needs sushi and a tiny fork.\n\nIt sends easy jobs to cheaper AI. You meet it in routing and company AI apps.", "humanExplainDisplay": "Cost-aware AI is the parent packing ==school lunch==.\nNot every Tuesday needs ==sushi and a tiny fork==.\n\nIt sends easy jobs to cheaper AI.\nYou meet it in routing and company AI apps.", "relationsNarrative": "Routing\nCost-aware AI uses Routing to save strong models for hard tasks.\n\nReasoning effort\nCost-aware AI raises or lowers Reasoning effort to balance cost and quality.\n\nLLMOps\nCost-aware AI is a common goal in LLMOps.\n\nAPI\nCost-aware AI changes API calls to avoid the pricey path.", "relations": { "model-routing": { "label": "drives … choices", "note": "It sends easy tasks to cheaper models and hard ones to stronger models." }, "reasoning-effort": { "label": "tunes …", "note": "It spends more thinking only when the task needs it." }, "llmops": { "label": "is part of …", "note": "Live AI systems must watch both cost and results." }, "api": { "label": "shapes … calls", "note": "Each API call has a price and a wait time." } } }, "zh": { "fullName": "成本感知 AI", "factExplain": "会把成本约束纳入决策与调用的 AI 用法。", "humanExplain": "像菜市场会过日子的阿姨：炖汤不必次次海参鲍鱼，青菜豆腐配得好，照样把一桌撑起来。\n\n常用于模型调度和企业部署，在效果可接受前提下尽量省钱。", "humanExplainDisplay": "像菜市场会过日子的阿姨：\n炖汤不必次次==海参鲍鱼==，\n青菜豆腐配得好，\n照样把一桌==撑起来==。\n\n常用于模型调度和企业部署，\n在效果可接受前提下尽量省钱。", "relationsNarrative": "Model-routing\n它常通过模型路由，把贵模型留给更难的任务。\n\nReasoning-effort\n它会调低或调高推理强度，交换效果与成本。\n\nLLMOps\n它是 LLMOps 里常见的优化目标之一。\n\nAPI\n它会影响接口调用策略，避免每次都走最贵方案。", "relations": { "model-routing": { "label": "驱动…分流", "note": "按任务难度选不同模型路线。" }, "reasoning-effort": { "label": "调节…强度", "note": "控制想多想少来换成本。" }, "llmops": { "label": "属于…关注点", "note": "线上部署常要盯成本与效果。" }, "api": { "label": "影响…调用", "note": "每次请求都要算钱和延迟。" } } } } }, { "id": "cross-entropy-loss", "name": "Cross-Entropy Loss", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "softmax" }, { "to": "gradient-descent" }, { "to": "backpropagation" }, { "to": "language-modeling" } ], "track": "history", "i18n": { "en": { "fullName": "Cross-Entropy Loss", "factExplain": "A loss function that scores how far predicted probabilities are from the truth.", "humanExplain": "Cross-entropy is a picky quiz teacher. Pick pizza for lunch, but mumble it, and you still lose points.\n\nIt scores how far the model’s probabilities are from the right answer. You meet it in classifiers and language models.", "humanExplainDisplay": "Cross-entropy is a ==picky quiz teacher==.\nPick ==pizza for lunch==,\nbut mumble it,\nand you still lose points.\n\nIt scores how far the model’s probabilities are\nfrom the right answer.\nYou meet it in classifiers\nand language models.", "relationsNarrative": "Softmax\nCross-Entropy often follows Softmax and scores its probabilities.\n\nGradient Descent\nCross-Entropy gives Gradient Descent a target to improve.\n\nBackpropagation\nCross-Entropy creates an error signal, and Backprop sends it back.\n\nLM\nLMs often use Cross-Entropy to learn the next word.", "relations": { "softmax": { "label": "often follows …", "note": "Softmax gives probabilities for Cross-Entropy to score." }, "gradient-descent": { "label": "guides …", "note": "A clear loss gives Gradient Descent a direction." }, "backpropagation": { "label": "sends errors through …", "note": "Backprop carries the error signal through the network." }, "language-modeling": { "label": "often trains …", "note": "LMs often use it to learn the next word." } } }, "zh": { "fullName": "交叉熵损失", "factExplain": "衡量模型预测分布与真实分布差距的损失函数。", "humanExplain": "像相亲时你不光得押中对方喜好，还得押得准；明明猜对方向却拿不准劲儿，照样减分。\n\n常用于分类和语言模型训练，衡量预测偏得有多离谱。", "humanExplainDisplay": "像相亲时你不光得==押中==对方喜好，\n还得押得==准==；\n明明猜对方向却拿不准劲儿，\n照样减分。\n\n常用于分类和语言模型训练，\n衡量预测偏得有多离谱。", "relationsNarrative": "Softmax\n它常接在 Softmax 后，对输出概率进行计分。\n\nGradient Descent\n它给梯度下降提供优化目标，指导参数更新。\n\nBackpropagation\n它产生的误差信号，会通过反向传播传回网络。\n\nLanguage Modeling\n语言模型预测下一个词时，常用它来训练。", "relations": { "softmax": { "label": "常接在…后", "note": "常把概率输出交给它来计分。" }, "gradient-descent": { "label": "给…提供方向", "note": "损失越清楚，参数更新越有方向。" }, "backpropagation": { "label": "通过…回传误差", "note": "误差信号靠反向传播传遍网络。" }, "language-modeling": { "label": "是…常用目标", "note": "预测下一个词时经常靠它训练。" } } } } }, { "id": "cross-validation", "name": "Cross-Validation", "layer": "L2", "era": "1970s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "benchmark-contamination" }, { "to": "hyperparameter-optimization" }, { "to": "bias-variance-tradeoff" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Cross-Validation", "factExplain": "A way to test a model by taking turns training and checking on data.", "humanExplain": "Cross-validation is like tasting soup from different spots in the pot. One spoonful may be all noodles and no broth.\n\nIt trains and checks the model on different data pieces, then rotates the pieces. You meet it with small datasets and model tuning.", "humanExplainDisplay": "Cross-validation is like tasting soup\nfrom ==different spots== in the pot.\nOne spoonful may be ==all noodles==\nand no broth.\n\nIt trains and checks the model\non different data pieces,\nthen rotates the pieces.\nYou meet it with small datasets\nand model tuning.", "relationsNarrative": "Benchmark contamination\nCross-validation reduces fake high scores from lucky test matches.\n\nHPO\nCross-validation helps compare different parameter settings.\n\nBias-Variance Tradeoff\nCross-validation helps show if a model overfits or underfits.\n\nSupervised Learning\nCross-validation is a common way to test supervised learning models.", "relations": { "benchmark-contamination": { "label": "helps spot … risk", "note": "Many test rounds reduce the chance of one lucky match." }, "hyperparameter-optimization": { "label": "often used in …", "note": "HPO often uses it to compare settings." }, "bias-variance-tradeoff": { "label": "helps watch …", "note": "It gives a steadier view of overfitting." }, "supervised-learning": { "label": "common in … evaluation", "note": "It is a common test method for classification and regression." } } }, "zh": { "fullName": "交叉验证", "factExplain": "一种用数据轮流训练与验证模型的评估方法。", "humanExplain": "网购别只看一条好评，得翻几页、换几家场景都看看；轮着验过，才知道它是真稳还是碰巧。\n\n常用于小数据集评估和调参，能更稳地估计模型真实表现。", "humanExplainDisplay": "网购别只看一条\n==好评==，\n得翻几页、换几家场景都看看；\n轮着验过，才知道它\n是真稳还是==碰巧==。\n\n常用于小数据集评估\n和调参，\n能更稳地估计模型真实表现。", "relationsNarrative": "Benchmark Contamination\n多轮验证能减少评测因撞题带来的虚高表现。\n\nHyperparameter Optimization\n它常被用来比较不同参数设置的效果。\n\nBias-Variance Tradeoff\n它能更稳地观察模型是否过拟合或欠拟合。\n\nSupervised Learning\n它是监督学习里常见的模型评估方法。", "relations": { "benchmark-contamination": { "label": "帮发现…风险", "note": "多轮验证能减少偶然撞题的错觉。" }, "hyperparameter-optimization": { "label": "常用于…调参", "note": "调参数时常拿它比较方案优劣。" }, "bias-variance-tradeoff": { "label": "用来观察…", "note": "能更稳地看模型是否过拟合。" }, "supervised-learning": { "label": "常见于…评估", "note": "分类回归任务里最常用的评估法。" } } } } }, { "id": "cuda", "name": "CUDA", "layer": "L5", "sublayer": "compute", "era": "2007", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "pytorch" }, { "to": "llm-inference-engine" }, { "to": "flash-attention" } ], "track": "history", "i18n": { "en": { "fullName": "Compute Unified Device Architecture", "factExplain": "NVIDIA’s toolset that lets programs use GPUs for many math jobs at once.", "humanExplain": "CUDA gives your graphics card a work badge. It stops drawing dragons and shows up in a hard hat to do math.\n\nIt lets programs send huge math jobs to the GPU. It sits under AI training and fast inference.", "humanExplainDisplay": "CUDA gives your graphics card a ==work badge==.\nIt stops drawing dragons\nand shows up in a ==hard hat== to do math.\n\nIt lets programs send huge math jobs\nto the GPU.\nIt sits under AI training\nand fast inference.", "relationsNarrative": "GPU\nCUDA turns the GPU from a graphics chip into a programmable math worker.\n\nPyTorch\nPyTorch often uses CUDA to call the GPU for deep learning work.\n\nInference engine\nAn inference engine often uses CUDA to run fast low-level math.\n\nFlash Attention\nFlash Attention is often built on CUDA for high speed.", "relations": { "gpu": { "label": "puts … to work", "note": "CUDA turns a GPU into a general math machine." }, "pytorch": { "label": "sits under …", "note": "PyTorch often uses CUDA to reach the GPU." }, "llm-inference-engine": { "label": "speeds up …", "note": "Inference engines use CUDA to run fast low-level math." }, "flash-attention": { "label": "runs …", "note": "Flash Attention is often built with CUDA for speed." } } }, "zh": { "fullName": "统一计算设备架构", "factExplain": "NVIDIA 提供的 GPU 通用并行计算平台与编程接口。", "humanExplain": "CUDA 像给显卡发了工牌：本来在网吧打游戏，现在进机房拧螺丝、搬大件、连夜赶工。\n\n它让程序把计算丢给 GPU 并行处理，是训练和推理加速的底层通道。", "humanExplainDisplay": "CUDA 像给显卡发了==工牌==：\n本来在网吧打游戏，\n现在进机房==连夜赶工==。\n\n它让程序把计算丢给 GPU 并行处理，\n是训练和推理加速的底层通道。", "relationsNarrative": "GPU\nCUDA 把 GPU 从图形芯片变成可编程计算设备。\n\nPytorch\n很多深度学习框架通过 CUDA 调用显卡算力。\n\nInference engine\n推理引擎常依赖 CUDA 执行底层高性能计算。\n\nFlash Attention\n这类高性能算子通常需要在 CUDA 上实现。", "relations": { "gpu": { "label": "让…干活", "note": "它把 GPU 变成通用计算工具。" }, "pytorch": { "label": "给…当底座", "note": "很多深度学习框架靠它调用显卡。" }, "llm-inference-engine": { "label": "支撑…加速", "note": "推理引擎常靠它驱动底层算子。" }, "flash-attention": { "label": "承载…实现", "note": "高性能算子通常要在它上面写。" } } } } }, { "id": "cudnn", "name": "CuDNN", "layer": "L5", "sublayer": "compute", "era": "2014", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "cuda" }, { "to": "pytorch" }, { "to": "cnn" }, { "to": "gpu" } ], "track": "history", "i18n": { "en": { "fullName": "CUDA Deep Neural Network Library", "factExplain": "NVIDIA’s low-level GPU library for speeding up deep learning.", "humanExplain": "cuDNN is a pit crew for deep learning. When the model pulls in, the heavy tire-change jobs happen fast.\n\nIt speeds up training and prediction under the hood. You often meet it hidden inside PyTorch and vision models.", "humanExplainDisplay": "cuDNN is a ==pit crew== for deep learning.\nWhen the model pulls in,\nthe ==heavy tire-change jobs== happen fast.\n\nIt speeds up training and prediction\nunder the hood.\nYou often meet it hidden inside\nPyTorch and vision models.", "relationsNarrative": "CUDA\ncuDNN is built on CUDA and uses it to run deep learning work faster.\n\nPyTorch\nPyTorch often calls cuDNN under the hood for GPU speed.\n\nCNN\ncuDNN speeds up the convolution and activation steps in CNNs.\n\nGPU\ncuDNN needs a GPU, because it helps the GPU handle heavy work.", "relations": { "cuda": { "label": "built on …", "note": "It uses CUDA to run work on the GPU." }, "pytorch": { "label": "is called by …", "note": "Many PyTorch operations use cuDNN for speed." }, "cnn": { "label": "speeds up … core ops", "note": "It gives CNNs fast convolutions and activations." }, "gpu": { "label": "needs … for speed", "note": "Without a GPU, cuDNN has little to show off." } } }, "zh": { "fullName": "CUDA 深度神经网络库", "factExplain": "NVIDIA 为深度学习提供的 GPU 加速底层库。", "humanExplain": "它像汽修厂备好的专用套筒：卷积、激活这些重活都顺手趁手，模型一上工位就能拧得飞快。\n\n常在底层给训练和推理提速，广泛藏在深度学习框架和视觉模型下面。", "humanExplainDisplay": "它像汽修厂\n备好的==专用套筒==：\n卷积、激活这些重活\n都顺手趁手，\n模型一上工位\n就能==拧得飞快==。\n\n常在底层给训练和推理提速，\n广泛藏在深度学习框架\n和视觉模型下面。", "relationsNarrative": "CUDA\n它建立在 CUDA 之上，负责把常见深度学习算子跑得更快。\n\nPyTorch\n很多 PyTorch 的底层计算，会调用它来拿到 GPU 加速。\n\nCNN\n卷积网络里的卷积和激活等操作，正是它最擅长优化的部分。\n\nGPU\n它要靠 GPU 才能发挥价值，本质是在替 GPU 安排重活。", "relations": { "cuda": { "label": "建立在…之上", "note": "它调用 CUDA 在 GPU 上干活。" }, "pytorch": { "label": "被…在底层调用", "note": "很多 PyTorch 算子会走它加速。" }, "cnn": { "label": "加速…核心算子", "note": "卷积网络尤其依赖它的优化实现。" }, "gpu": { "label": "依赖…发挥性能", "note": "没有 GPU，它就没法大显身手。" } } } } }, { "id": "curse-of-dimensionality", "name": "Curse of Dimensionality", "layer": "L2", "era": "1957", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "dimensionality-reduction" }, { "to": "embedding" }, { "to": "vector-search" }, { "to": "feature-selection" } ], "track": "history", "i18n": { "en": { "fullName": "Curse of Dimensionality", "factExplain": "With many dimensions, data gets sparse, and distance and computing get harder.", "humanExplain": "Dim Curse is like looking for your keys in a mansion. After room 80, every couch looks equally guilty.\n\nIn vector search and clustering, too many dimensions make distances blur. Models with lots of features feel it too.", "humanExplainDisplay": "Dim Curse is like\nlooking for your keys in a ==mansion==.\nAfter room 80,\nevery couch looks ==equally guilty==.\n\nIn vector search and clustering,\ntoo many dimensions make distances blur.\nModels with lots of features feel it too.", "relationsNarrative": "Dim. Reduction\nDim. Reduction is one common way to ease Dim Curse.\n\nEmbedding\nWhen embeddings have too many dimensions, distance gets less useful.\n\nVector search\nDim Curse makes nearest-neighbor search harder and less reliable.\n\nFeature Selection\nFeature Selection removes useless features, so the space gets less sparse.", "relations": { "dimensionality-reduction": { "label": "pushes use of …", "note": "Dim. Reduction is a common way to ease sparse high-dimensional data." }, "embedding": { "label": "warps … space", "note": "Higher-dimensional embeddings can make distances less trustworthy." }, "vector-search": { "label": "makes … harder", "note": "In high dimensions, the nearest match is harder to trust." }, "feature-selection": { "label": "needs …", "note": "Removing useless features can ease Dim Curse." } } }, "zh": { "fullName": "维度灾难", "factExplain": "维度一高，数据会更稀疏，距离和计算都更难用。", "humanExplain": "夜市点麻辣烫时，配料越勾越细，勺子捞半天像在清汤里找针：看着选择多，实际谁都离得差不多。\n\n它会让距离失真，常见于检索、聚类和高维建模。", "humanExplainDisplay": "夜市点麻辣烫时，\n配料越勾越细，\n勺子捞半天像在清汤里\n找==针==：看着选择多，\n实际谁都==离得差不多==。\n\n它会让距离失真，\n常见于检索、聚类\n和高维建模。", "relationsNarrative": "Dimensionality Reduction\n降维是缓解它最常见的办法之一。\n\nEmbedding\n向量表示一旦维度太高，距离就没那么好用了。\n\nVector Search\n它会让近邻检索变难，搜索效率和效果都受影响。\n\nFeature Selection\n筛掉无关特征，能减少高维带来的稀疏和噪声。", "relations": { "dimensionality-reduction": { "label": "推动…降维", "note": "降维常用来缓解高维稀疏问题。" }, "embedding": { "label": "影响…空间", "note": "向量维度越高，距离越容易失真。" }, "vector-search": { "label": "增加…难度", "note": "高维空间会削弱近邻检索直觉。" }, "feature-selection": { "label": "需要…筛特征", "note": "删掉无关维度能减轻维度灾难。" } } } } }, { "id": "cursor", "name": "Cursor", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-05-23T09:50:00Z", "relations": [ { "to": "copilot" }, { "to": "vibe-coding" }, { "to": "agent" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Cursor", "factExplain": "A coding tool with AI built deep into the programming workflow.", "humanExplain": "Cursor is like a coding buddy in the next chair. It types fast. Sometimes it breaks your code with a straight face.\n\nIt helps with bugs, components, and codebases. Coding starts to feel less like typing and more like giving directions.", "humanExplainDisplay": "Cursor is like a ==coding buddy==\nin the next chair.\nIt types fast.\nSometimes it breaks your code\nwith a ==straight face==.\n\nIt helps with bugs, components,\nand codebases.\nCoding starts to feel less like typing\nand more like giving directions.", "relationsNarrative": "Copilot\nCursor continues Copilot-style coding help and brings it deeper into the IDE.\n\nVibe-coding\nIn Cursor, Vibe-coding can turn instructions into code fast.\n\nAgent\nCursor can act like an Agent when it handles multi-step coding tasks.\n\nLLM\nCursor mainly uses an LLM to read, change, and explain code.", "relations": { "copilot": { "label": "is in the same family as …", "note": "Cursor builds on Copilot-style coding help inside the editor." }, "vibe-coding": { "label": "helps drive …", "note": "Cursor makes it easier to turn plain instructions into code." }, "agent": { "label": "can act like an …", "note": "Cursor can show Agent behavior during multi-step coding work." }, "llm": { "label": "relies on …", "note": "LLMs help Cursor read, change, and explain code." } } }, "zh": { "fullName": "Cursor", "factExplain": "一款把 AI 深度集成进编码流程的开发工具。", "humanExplain": "Cursor 像坐在你旁边的代码同事，能补代码、读项目，也会偶尔一本正经地改坏。\n\n它适合改 bug、写组件、理解代码库，让编程从敲字变成指挥。", "humanExplainDisplay": "Cursor 像坐在你旁边的\n==代码同事==。\n能补代码、读项目，\n也会偶尔一本正经地改坏。\n\n它适合改 bug、写组件、\n理解代码库。\n编程从敲字，\n慢慢变成指挥。", "relationsNarrative": "Copilot\nCursor 延续了 Copilot 的编程辅助能力，并进一步集成到 IDE。\n\nVibe-coding\nVibe-coding 在 Cursor 中更容易从指令直接落到代码。\n\nAgent\nCursor 执行多步骤开发任务时，会呈现 Agent 特征。\n\nLLM\nCursor 的代码理解、修改和解释能力主要来自 LLM。", "relations": { "copilot": { "label": "与…同类" }, "vibe-coding": { "label": "推动…" }, "llm": { "label": "依赖…辅助" } } } } }, { "id": "custom-ai-chip", "name": "Custom AI Chip", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "ai-chip" }, { "to": "gpu" }, { "to": "tpu" }, { "to": "ai-data-center" } ], "track": "today", "i18n": { "en": { "fullName": "Custom AI Chip", "factExplain": "A computer chip built for one kind of AI work.", "humanExplain": "A custom AI chip is like a waffle maker. Bad at soup, great at waffles.\n\nYou meet it in data centers and some devices. It helps AI train and answer with less power and cost.", "humanExplainDisplay": "A custom AI chip is like a ==waffle maker==.\nBad at soup,\ngreat at ==waffles==.\n\nYou meet it in data centers\nand some devices.\nIt helps AI train and answer\nwith less power and cost.", "relationsNarrative": "AI chip\nA custom AI chip is the focused path inside AI chips.\n\nGPU\nIt is more focused than a GPU, so it can cut cost and power use.\n\nTPU\nA TPU is a famous custom chip for AI work.\n\nAI data center\nAn AI data center uses it to save power and lower answer costs.", "relations": { "ai-chip": { "label": "specializes …", "note": "It is a special type of AI chip." }, "gpu": { "label": "replaces or helps …", "note": "A GPU is more general. A custom chip is more focused." }, "tpu": { "label": "has … as a famous example", "note": "A TPU is a classic custom AI chip." }, "ai-data-center": { "label": "serves …", "note": "Data centers use it to lower AI answer costs." } } }, "zh": { "fullName": "定制 AI 芯片", "factExplain": "为特定 AI 负载专门设计的计算芯片。", "humanExplain": "定制 AI 芯片像短跑钉鞋：走路硌脚不万能，冲刺那几秒把通用球鞋甩开。\n\n用于数据中心和端侧，让训练推理更省电、更便宜。", "humanExplainDisplay": "定制 AI 芯片像==短跑钉鞋==：\n走路硌脚不万能，\n冲刺那几秒\n把==通用球鞋==甩开。\n\n用于数据中心和端侧，\n让训练推理更省电、\n更便宜。", "relationsNarrative": "AI chip\n定制芯片是 AI 芯片的专用化路线。\n\nGPU\n它比 GPU 更专精，常用于降成本和功耗。\n\nTPU\nTPU 是面向 AI 负载定制芯片的代表。\n\nAI data center\n数据中心用它提升能效、压低推理成本。", "relations": { "ai-chip": { "label": "专门化…", "note": "它是 AI 芯片的专用化分支。" }, "gpu": { "label": "替代或补充…", "note": "GPU 更通用，定制芯片更专精。" }, "tpu": { "label": "以…为代表", "note": "TPU 是典型的定制 AI 芯片。" }, "ai-data-center": { "label": "服务…", "note": "数据中心用它压低推理成本。" } } } } }, { "id": "cyc", "name": "Cyc", "layer": "L5", "sublayer": "product", "era": "1984", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "commonsense-reasoning" }, { "to": "symbolic-ai" }, { "to": "ontology" } ], "track": "history", "i18n": { "en": { "fullName": "Cyc", "factExplain": "A large hand-built knowledge base full of everyday common sense.", "humanExplain": "Cyc is like a fussy school librarian for real life. It files cards like, “Cups can fall,” and “Dogs are not sandwiches.”\n\nIt helps machines use common sense and organize knowledge. It also shows how hard hand-written knowledge can get.", "humanExplainDisplay": "Cyc is like a ==fussy school librarian== for real life.\nIt files cards like,\n==“Cups can fall,”==\nand “Dogs are not sandwiches.”\n\nIt helps machines use common sense\nand organize knowledge.\nIt also shows how hard\nhand-written knowledge can get.", "relationsNarrative": "KR\nCyc uses KR to write common sense in a machine-readable form.\n\nCommonsense Reasoning\nCyc aims to give machines everyday common sense.\n\nSymbolic AI\nCyc is a classic Symbolic AI project based on hand-written knowledge.\n\nOntology\nCyc uses Ontology to define concepts, types, and relations.", "relations": { "knowledge-representation": { "label": "organizes common sense with …", "note": "Cyc writes common sense in a form machines can reason with." }, "commonsense-reasoning": { "label": "serves …", "note": "Cyc aims to help machines reason about everyday life." }, "symbolic-ai": { "label": "belongs to …", "note": "Cyc is a classic hand-built knowledge project in Symbolic AI." }, "ontology": { "label": "builds on …", "note": "Cyc uses Ontology to define concepts and relations." } } }, "zh": { "fullName": "常识知识库项目", "factExplain": "一个手工构建的大规模常识知识库。", "humanExplain": "Cyc 像给 AI 请来居委会大妈：杯子会掉地上这种小常识都登记。\n\n用于常识推理和知识表示，也暴露手写知识的难。", "humanExplainDisplay": "Cyc 像给 AI 请来\n==居委会大妈==：\n杯子会掉地上这种\n小常识都==登记==。\n\n用于常识推理和知识表示，\n也暴露手写知识的难。", "relationsNarrative": "Knowledge Representation\nCyc 用知识表示把常识写成机器可推理的形式。\n\nCommonsense Reasoning\nCyc 的目标是让机器具备日常常识推理。\n\nSymbolic AI\nCyc 是符号 AI 手工编码知识路线的代表。\n\nOntology\nCyc 依赖本体来定义概念、类别和关系。", "relations": { "knowledge-representation": { "label": "用…组织常识", "note": "把常识写成可推理的表示。" }, "commonsense-reasoning": { "label": "服务…", "note": "目标是让机器会日常推理。" }, "symbolic-ai": { "label": "属于…路线", "note": "手工造知识的经典代表。" }, "ontology": { "label": "构建…", "note": "用本体定义概念和关系。" } } } } }, { "id": "dartmouth-workshop", "name": "Dartmouth Workshop", "layer": "L1", "era": "1956", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "turing-test" }, { "to": "agi" }, { "to": "neural-network" } ], "track": "history", "i18n": { "en": { "fullName": "Dartmouth Conference", "factExplain": "A 1956 meeting where researchers first gave “artificial intelligence” its official name.", "humanExplain": "Dartmouth was AI’s big name tag. It put the brainy lunch-table crowd under one club-room sign.\n\nIn 1956, this meeting gave the field its name. It helped AI become an organized research field.", "humanExplainDisplay": "Dartmouth was AI’s ==big name tag==.\nIt put the brainy lunch-table crowd\nunder one ==club-room sign==.\n\nIn 1956, this meeting gave the field its name.\nIt helped AI become an organized research field.", "relationsNarrative": "Turing-test\nDartmouth built on early machine intelligence ideas and pushed them into a field.\n\nAGI\nThe dream of general intelligence was already there, and later people called it AGI.\n\nNeural-network\nDartmouth came before today’s neural-network boom and marks an earlier AI starting point.", "relations": { "turing-test": { "label": "builds on early ideas from …", "note": "It turned early talk about machine intelligence into a research field." }, "agi": { "label": "planted the goal of …", "note": "The dream of general intelligence was there from the start." }, "neural-network": { "label": "came before the rise of …", "note": "It was an early AI starting point before today’s neural-network boom." } } }, "zh": { "fullName": "达特茅斯会议", "factExplain": "1956 年首次正式提出“人工智能”之名的学术会议。", "humanExplain": "像给一群散兵游勇正式挂了块招牌：以前各打各的，从这回起大家有了统一门头。\n\n它标志着 AI 被正式命名，并开始成为一门组织化研究领域。", "humanExplainDisplay": "像给一群散兵游勇\n正式挂了块==招牌==：\n以前各打各的，\n从这回起大家有了统一==门头==。\n\n它标志着 AI 被正式命名，\n并开始成为一门组织化研究领域。", "relationsNarrative": "Turing-test\n它承接更早的机器智能讨论，并把问题推进到学科层面。\n\nAGI\n通用智能的雄心在那时就已出现，后来被称作 AGI。\n\nNeural-network\n它发生在当代神经网络崛起前，是更早期的 AI 起点。", "relations": { "turing-test": { "label": "承接…早期思想", "note": "它把更早的机器智能讨论推向学科化。" }, "agi": { "label": "为…埋下目标", "note": "通用智能理想从早期就已存在。" }, "neural-network": { "label": "属于…早期背景", "note": "它早于当代神经网络浪潮与爆发。" } } } } }, { "id": "data-augmentation", "name": "Data Augmentation", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "computer-vision" }, { "to": "imagenet" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Data Augmentation", "factExplain": "A way to grow training data by changing the examples you already have.", "humanExplain": "Take one dog photo and give it a little wobble. Now the AI learns “dog,” not one perfect puppy pose.\n\nIt is big in image training, and shows up in speech and text. With few examples, it helps models handle new cases better.", "humanExplainDisplay": "Take one dog photo\nand give it a ==little wobble==.\nNow the AI learns “dog,”\nnot ==one perfect puppy pose==.\n\nIt is big in image training,\nand shows up in speech and text.\nWith few examples,\nit helps models handle new cases better.", "relationsNarrative": "Regularization\nData Augmentation is a common way to reduce overfitting.\n\nComputer Vision\nIt is most common and most mature in image training.\n\nImageNet\nLarge ImageNet training often treats it as a basic step.\n\nSupervised Learning\nIt gives labeled training more varied examples.", "relations": { "regularization": { "label": "is a common form of …", "note": "It adds changed examples to reduce overfitting." }, "computer-vision": { "label": "is often used in … training", "note": "Flips and crops became common early in vision work." }, "imagenet": { "label": "is routine in … training", "note": "ImageNet training often uses it as a basic step." }, "supervised-learning": { "label": "often serves …", "note": "Labeled training uses it when examples are too few." } } }, "zh": { "fullName": "数据增强（Data Augmentation）", "factExplain": "通过变换训练数据来扩充样本的方法。", "humanExplain": "老师要是总拿原题考，学霸也可能只会背答案；换个问法、倒着问，才真测出会没会。\n\n常用于图像、语音、文本训练，样本少时尤其有用，可提升泛化。", "humanExplainDisplay": "老师要是总拿\n原题考，\n学霸也可能\n只会==背答案==；\n换个问法、倒着问，\n才真测出==会没会==。\n\n常用于图像、语音、文本训练，\n样本少时尤其有用，\n可提升泛化。", "relationsNarrative": "Regularization\n它是减少过拟合的常见办法之一。\n\nComputer Vision\n它在图像训练里最常见，也最成熟。\n\nImageNet\n大规模视觉训练常把它当基本操作。\n\nSupervised Learning\n它常用来给有标签训练补样本变化。", "relations": { "regularization": { "label": "属于…常见手段", "note": "它常靠扩充变化来减少过拟合。" }, "computer-vision": { "label": "常用于…训练", "note": "翻转裁剪扰动最早在视觉里最常见。" }, "imagenet": { "label": "常配合…使用", "note": "大规模视觉训练常把它当基本操作。" }, "supervised-learning": { "label": "常服务于…", "note": "有标签训练最常拿它补样本不足。" } } } } }, { "id": "data-donation", "name": "Data Donation", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "differential-privacy" }, { "to": "copyright" } ], "track": "today", "i18n": { "en": { "fullName": "Data Donation", "factExplain": "The free sharing of personal data for AI training or research.", "humanExplain": "Data donation is like letting a researcher peek in your junk drawer. It helps science, but it may show your weird receipts.\n\nIt is used for training data and public research. It helps models understand real people better.", "humanExplainDisplay": "Data donation is like letting a researcher\npeek in your ==junk drawer==.\nIt helps science,\nbut it may show your ==weird receipts==.\n\nIt is used for training data\nand public research.\nIt helps models understand\nreal people better.", "relationsNarrative": "Data-privacy\nData donation uses real life data, so privacy must be handled first.\n\nDifferential Privacy\nDifferential Privacy can protect data before it enters training.\n\nCopyright\nDonated content can include works, so ownership may matter.", "relations": { "data-privacy": { "label": "can expose …", "note": "The more real the data is, the higher the privacy risk." }, "differential-privacy": { "label": "can lower risk with …", "note": "Differential Privacy hides personal clues before training." }, "copyright": { "label": "can overlap with …", "note": "Photos, text, and voice may have owners." } } }, "zh": { "fullName": "数据捐赠", "factExplain": "用户自愿把个人数据提供给 AI 训练或研究。", "humanExplain": "这事像去医院献血：捐出去的不是钱，是你身上的“样本”，能救研究，也最怕信息泄露。\n\n常用于训练数据收集和公益研究，让模型更懂真人情况。", "humanExplainDisplay": "这事像去医院==献血==：\n捐出去的不是钱，\n是你身上的==“样本”==，\n能救研究，也最怕信息泄露。\n\n常用于训练数据收集\n和公益研究，\n让模型更懂真人情况。", "relationsNarrative": "Data-privacy\n数据捐赠越接近真实生活，隐私风险越需要先处理。\n\nDifferential Privacy\n差分隐私可先做保护，再降低数据进入训练的风险。\n\nCopyright\n被捐出的内容若含作品，也可能涉及版权归属。", "relations": { "data-privacy": { "label": "容易碰到…", "note": "捐赠越真实，隐私风险越高。" }, "differential-privacy": { "label": "可用…降风险", "note": "先做匿名保护，再进入训练流程。" }, "copyright": { "label": "常与…交叉", "note": "照片文字语音都可能牵涉权属。" } } } } }, { "id": "data-exfiltration", "name": "Exfiltration", "layer": "L6", "era": "2010s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "prompt-injection" }, { "to": "permission-fatigue" }, { "to": "ai-sandbox" } ], "track": "today", "i18n": { "en": { "fullName": "Data exfiltration", "factExplain": "The unauthorized removal of data out of a system.", "humanExplain": "It is not a nosy coworker reading one file. It is them wheeling out the whole filing cabinet.\n\nIt means private data leaves the system. It can start with a stolen account. It can also come from a shady plug-in.", "humanExplainDisplay": "It is not a nosy coworker\nreading one file.\nIt is them ==wheeling out==\nthe whole ==filing cabinet==.\n\nIt means private data leaves the system.\nIt can start with a stolen account.\nIt can also come from a shady plug-in.", "relationsNarrative": "Data-privacy\nData exfiltration is a classic data privacy failure.\n\nPrompt injection\nPrompt injection can trick AI into showing and sending private data.\n\nPermission fatigue\nPermission fatigue makes users grant access they should not give.\n\nAI sandbox\nAn AI sandbox can reduce the chance data gets carried away.", "relations": { "data-privacy": { "label": "is a … incident", "note": "Data exfiltration can break privacy rules and laws." }, "prompt-injection": { "label": "can be triggered by …", "note": "A malicious prompt can trick AI into sending out secrets." }, "permission-fatigue": { "label": "often slips through …", "note": "When users tap Allow too fast, leaks become more likely." }, "ai-sandbox": { "label": "is limited by …", "note": "A sandbox can limit what data a tool can reach or send." } } }, "zh": { "fullName": "Data exfiltration／数据外泄", "factExplain": "指数据被未经授权地导出或带走。", "humanExplain": "数据外泄像把公司机密装进外卖袋，门禁没响，人已经下楼。\n\nAI 应用里，它会带走密钥和客户资料，尤其在代理、插件场景。", "humanExplainDisplay": "数据外泄像把公司机密\n装进==外卖袋==，\n==门禁没响==，\n人已经下楼。\n\nAI 应用里，它会带走\n密钥和客户资料，\n尤其在代理、插件场景。", "relationsNarrative": "Data-privacy\n数据外泄是数据隐私受损的典型后果之一。\n\nPrompt injection\n恶意提示可能诱导 AI 暴露并导出敏感数据。\n\nPermission fatigue\n用户疲于确认授权时，更容易放出不该给的权限。\n\nAI sandbox\n沙箱通过隔离环境，减少数据被带走的机会。", "relations": { "data-privacy": { "label": "属于…事故", "note": "数据外泄是隐私与合规风险的典型场景。" }, "prompt-injection": { "label": "可被…诱发", "note": "恶意提示可能诱导系统泄露机密内容。" }, "permission-fatigue": { "label": "常借…得手", "note": "用户乱点授权时，外泄风险会明显上升。" }, "ai-sandbox": { "label": "靠…隔离风险", "note": "沙箱可限制数据访问与外发范围。" } } } } }, { "id": "data-labeling", "name": "Data Labeling", "layer": "L2", "era": "2000s", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "human-in-the-loop" }, { "to": "data-augmentation" }, { "to": "imagenet" } ], "track": "today", "i18n": { "en": { "fullName": "Data Labeling", "factExplain": "The process of adding answers or categories to training data.", "humanExplain": "Data labeling is like putting stickers on school lunch trays. Pizza gets a pizza sticker, not a mystery slime sticker.\n\nIt powers supervised training, tests, and fine-tuning. Bad labels teach the model the wrong lesson.", "humanExplainDisplay": "Data labeling is like putting ==stickers==\non school lunch trays.\nPizza gets a pizza sticker,\nnot a ==mystery slime== sticker.\n\nIt powers supervised training, tests, and fine-tuning.\nBad labels teach the model the wrong lesson.", "relationsNarrative": "Supervised Learning\nData Labeling gives Supervised Learning the answer key for each example.\n\nHuman-in-the-loop\nHuman-in-the-loop helps create, check, and fix labels.\n\nData Augmentation\nData Augmentation can make more training data from labeled examples.\n\nImageNet\nImageNet became powerful because people labeled a huge number of images.", "relations": { "supervised-learning": { "label": "gives answers to …", "note": "Supervised Learning needs examples with labels." }, "human-in-the-loop": { "label": "depends on … for checks", "note": "People help make labels clear and correct." }, "data-augmentation": { "label": "works with … to grow data", "note": "Labeled examples are often expanded with Data Augmentation." }, "imagenet": { "label": "helped build …", "note": "ImageNet became famous through huge human labeling work." } } }, "zh": { "fullName": "数据标注", "factExplain": "为训练数据添加答案或类别的过程。", "humanExplain": "数据标注像给模型的练习册配标准答案：这页是「苹果」，那页是「猫」。\n\n它支撑监督训练、评测和微调，标错会教歪模型。", "humanExplainDisplay": "数据标注像给模型的练习册\n配==标准答案==：\n这页是「苹果」，\n那页是==「猫」==。\n\n它支撑监督训练、评测，\n和微调，\n标错会教歪模型。", "relationsNarrative": "Supervised Learning\n数据标注为监督学习提供可对照的标准答案。\n\nHuman-in-the-loop\n人工参与常用于制定、审核和修正标签。\n\nData Augmentation\n标注后的样本可再增强，扩大训练数据规模。\n\nImageNet\nImageNet 的影响力建立在大规模人工标注上。", "relations": { "supervised-learning": { "label": "为…提供答案", "note": "监督学习离不开带标签样本。" }, "human-in-the-loop": { "label": "依赖…把关", "note": "人工参与决定标签质量。" }, "data-augmentation": { "label": "配合…扩充数据", "note": "标注样本常再被增强使用。" }, "imagenet": { "label": "支撑…成型", "note": "ImageNet 靠大规模标注出圈。" } } } } }, { "id": "data-parallelism", "name": "Data Parallelism", "layer": "L5", "sublayer": "architecture", "era": "2012", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "distributed-computing" }, { "to": "model-parallelism" }, { "to": "gpu" }, { "to": "sgd" } ], "track": "history", "i18n": { "en": { "fullName": "Data Parallelism", "factExplain": "A way to train copies of one model on many devices at once.", "humanExplain": "Data parallelism is like a cafeteria line with the same lasagna recipe. Each cook handles a different tray, then they compare notes before the next batch.\n\nIn AI training, each GPU gets a copy of the model and different data. It speeds up big model training, but syncing updates costs time.", "humanExplainDisplay": "Data parallelism is like a cafeteria line\nwith the ==same lasagna recipe==.\nEach cook handles ==a different tray==,\nthen they compare notes\nbefore the next batch.\n\nIn AI training,\neach GPU gets a copy of the model\nand different data.\nIt speeds up big model training,\nbut syncing updates costs time.", "relationsNarrative": "Distributed Computing\nData parallelism is one of the most common ways to do distributed training.\n\nModel parallelism\nData parallel splits the data, while model parallel splits the model.\n\nGPU\nData parallelism usually copies the same model onto many GPUs.\n\nSGD\nEach device computes gradients first, then they combine them for one update.", "relations": { "distributed-computing": { "label": "is part of … training", "note": "Data parallelism is one of the most common distributed training methods." }, "model-parallelism": { "label": "is often compared with …", "note": "Data parallel splits data; model parallel splits the model." }, "gpu": { "label": "hands work to …s", "note": "Each GPU usually holds its own copy of the same model." }, "sgd": { "label": "runs … in parallel", "note": "Each device computes gradients, then all devices update together." } } }, "zh": { "fullName": "数据并行", "factExplain": "把同一模型复制到多设备上并行训练的方法。", "humanExplain": "数据并行像同卷分班考试：每个考场做同一套题，最后把分数统一登总表。\n\n它让多卡并行处理不同样本提速，常见于大模型训练；但每轮同步参数会有通信开销。", "humanExplainDisplay": "数据并行像==同卷分班考试==：\n每个考场做同一套题，\n最后把分数统一登==总表==。\n\n它让多卡并行处理不同样本提速，\n常见于大模型训练；\n但每轮同步参数会有通信开销。", "relationsNarrative": "Distributed Computing\n数据并行是分布式训练里最常见的做法之一。\n\nModel parallelism\n它和模型并行常被放在一起比较：一个分数据，一个分模型。\n\nGPU\n数据并行通常把同一模型副本复制到多张 GPU 上。\n\nSGD\n各设备先分别算梯度，再汇总后做一次参数更新。", "relations": { "distributed-computing": { "label": "属于…训练套路", "note": "它是分布式训练里最常见的一类。" }, "model-parallelism": { "label": "常与…对比", "note": "一个分数据，一个分模型本体。" }, "gpu": { "label": "把任务分给…", "note": "通常把同一模型副本放到多张卡上。" }, "sgd": { "label": "并行执行…", "note": "各卡先各算梯度，再统一更新参数。" } } } } }, { "id": "data-privacy", "name": "Data-privacy", "layer": "L6", "era": "2023", "publishedAt": "2026-05-23T11:30:00Z", "relations": [ { "to": "local-llm" }, { "to": "api" }, { "to": "ai-regulation" }, { "to": "copyright" } ], "i18n": { "en": { "fullName": "Data Privacy", "factExplain": "Rules and risks around how AI collects, uses, and protects data.", "humanExplain": "Data privacy is like handing your phone to a nosy friend. You want help finding a photo, not a tour of your group chats.\n\nAt work, it means guarding private files and access. It also means checking what outside APIs keep.", "humanExplainDisplay": "Data privacy is like handing your phone to a ==nosy friend==.\nYou want help finding a photo,\nnot a tour of your ==group chats==.\n\nAt work, it means guarding private files and access.\nIt also means checking what outside APIs keep.", "relationsNarrative": "Local-LLM\nA Local-LLM can lower the risk of sensitive data leaving your system.\n\nAPI\nAPI calls add risk around data transfer, access, and storage.\n\nAI-regulation\nAI-regulation turns Data Privacy needs into legal duties.\n\nCopyright\nCopyright and Data Privacy both limit how data can be used.", "relations": { "local-llm": { "label": "can be protected by …", "note": "A Local-LLM can keep sensitive data inside your own system." }, "api": { "label": "can leak through …", "note": "API calls can send data outside your walls." }, "ai-regulation": { "label": "is governed by …", "note": "AI-regulation turns privacy rules into duties." }, "copyright": { "label": "overlaps with …", "note": "Copyright and data privacy both limit free use of data." } } }, "zh": { "fullName": "数据隐私", "factExplain": "个人或组织数据在 AI 使用中的收集、处理和保护问题。", "humanExplain": "数据隐私就是别让 AI 像家庭群大喇叭，把你的身份证号到处转发。\n\n它影响训练、问答和企业接入，决定哪些数据能被看见、记住和带走。", "humanExplainDisplay": "数据隐私就是别让 AI\n像==家庭群大喇叭==，\n把你的身份证号==到处转发==。\n\n它影响训练、问答和企业接入，\n决定哪些数据能被看见、记住和带走。", "relationsNarrative": "Local-LLM\nLocal-LLM 通过本地部署减少敏感数据外传风险。\n\nAPI\nAPI 调用会增加数据传输、权限和留存管理压力。\n\nAI-regulation\nAI-regulation 将 Data-privacy 要求转化为合规义务。\n\nCopyright\nCopyright 和 Data-privacy 都约束数据能否被任意使用。", "relations": { "local-llm": { "label": "可被…保护" }, "ai-regulation": { "label": "由…规范" }, "copyright": { "label": "关联…" } } } } }, { "id": "data-retention", "name": "Data retention", "layer": "L6", "era": "2010s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "ai-regulation" }, { "to": "llmops" }, { "to": "agent-memory" } ], "track": "today", "i18n": { "en": { "fullName": "Data retention", "factExplain": "The rules for how long data is kept and when it gets deleted.", "humanExplain": "Data retention is the fridge rule for leftovers. Keep pizza too long, and it grows a tiny science project.\n\nDelete it too soon, and nobody can check what happened later. You meet it in chat history, system logs, and training data.", "humanExplainDisplay": "Data retention is the ==fridge rule== for leftovers.\nKeep pizza too long,\nand it grows a ==tiny science project==.\n\nDelete it too soon,\nand nobody can check what happened later.\nYou meet it in chat history,\nsystem logs,\nand training data.", "relationsNarrative": "Data-privacy\nLonger data retention often means higher risk of leaks or misuse.\n\nAI-regulation\nMany compliance rules say how long data stays and how it gets deleted.\n\nLLMOps\nRetention rules for logs, chats, and training data are part of operations.\n\nMemory\nAn Agent’s long-term memory must know how long its data can stay.", "relations": { "data-privacy": { "label": "changes … risk", "note": "The longer data is kept, the larger the privacy risk often gets." }, "ai-regulation": { "label": "is limited by …", "note": "Many rules say how long data can stay and when it must go." }, "llmops": { "label": "belongs to … governance", "note": "Logs and data retention are part of running AI systems well." }, "agent-memory": { "label": "limits …", "note": "Long-term memory must first answer how long data may stay." } } }, "zh": { "fullName": "数据保留", "factExplain": "规定数据该存多久、何时删除的规则。", "humanExplain": "像宿舍衣柜换季：压箱底太久占地方还发霉，扔太早一降温又抓瞎，留多久最难拿捏。\n\n常见于聊天记录、日志和训练数据管理，影响隐私、合规与排障。", "humanExplainDisplay": "像宿舍衣柜换季：\n压箱底太久==占地方还发霉==，\n扔太早一降温又抓瞎，\n==留多久最难拿捏==。\n\n常见于聊天记录、\n日志和训练数据管理，\n影响隐私、合规与排障。", "relationsNarrative": "Data-privacy\n数据留存时间越长，隐私泄露和滥用风险通常越高。\n\nAI-regulation\n很多合规要求会直接规定数据该留多久、怎么删。\n\nLLMOps\n日志、对话和训练数据的留存策略属于运维治理。\n\nMemory\nAgent 的长期记忆设计，先要回答数据能留多久。", "relations": { "data-privacy": { "label": "影响…风险", "note": "留得越久，隐私暴露面通常越大。" }, "ai-regulation": { "label": "受…约束", "note": "很多法规会规定保存多久、何时删除。" }, "llmops": { "label": "属于…治理", "note": "日志与数据留存是运维治理的一部分." }, "agent-memory": { "label": "约束…记忆", "note": "长期记忆要先回答能留多久。" } } } } }, { "id": "dbscan", "name": "DBSCAN", "layer": "L2", "era": "1996", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "clustering" }, { "to": "unsupervised-learning" }, { "to": "k-means-clustering" }, { "to": "outlier-detection" } ], "track": "history", "i18n": { "en": { "fullName": "Density-Based Spatial Clustering of Applications with Noise", "factExplain": "An algorithm that groups crowded points and marks lonely points as noise.", "humanExplain": "DBSCAN reads data like a school dance map. Snack-table crowds are groups. One kid doing the robot alone is noise.\n\nIt finds weird-shaped groups in messy data. It can also flag lonely points as outliers.", "humanExplainDisplay": "DBSCAN reads data like a ==school dance map==.\nSnack-table crowds are groups.\nOne kid ==doing the robot alone== is noise.\n\nIt finds weird-shaped groups in messy data.\nIt can also flag lonely points as outliers.", "relationsNarrative": "Clustering\nDBSCAN groups nearby points by density.\n\nUnsupervised Learning\nDBSCAN finds groups without human labels.\n\nK-Means Clustering\nDBSCAN handles weird-shaped groups better than K-Means.\n\nOutlier Detection\nDBSCAN marks lonely low-density points as noise or outliers.", "relations": { "clustering": { "label": "is a kind of …", "note": "DBSCAN is a classic density-based clustering method." }, "unsupervised-learning": { "label": "works in …", "note": "It groups data without human labels." }, "k-means-clustering": { "label": "differs from …", "note": "It does not need groups to look like round blobs." }, "outlier-detection": { "label": "also does …", "note": "Lonely low-density points get marked as noise." } } }, "zh": { "fullName": "基于密度的空间聚类算法", "factExplain": "按样本密度聚类并标出噪声点的算法。", "humanExplain": "DBSCAN 像演唱会看人浪：挤成片的算一团，落单自拍的当噪声。\n\n用于不规则聚类，也能顺手标出异常点。", "humanExplainDisplay": "DBSCAN 像==演唱会看人浪==：\n挤成片的算一团，\n==落单自拍的当噪声==。\n\n用于不规则聚类，\n也能顺手，\n标出异常点。", "relationsNarrative": "Clustering\nDBSCAN 是按密度把相近样本分成簇的方法。\n\nUnsupervised Learning\nDBSCAN 不需要人工标签，也能自动发现分组。\n\nK-Means Clustering\nDBSCAN 比 K-Means 更能处理不规则形状的簇。\n\nOutlier Detection\nDBSCAN 会把低密度孤点标成噪声或异常。", "relations": { "clustering": { "label": "属于…方法", "note": "它是经典的密度聚类算法。" }, "unsupervised-learning": { "label": "用于…", "note": "它不用标签也能自动分组。" }, "k-means-clustering": { "label": "对比…", "note": "它不要求簇必须像圆团。" }, "outlier-detection": { "label": "顺手做…", "note": "低密度孤点会被标为噪声。" } } } } }, { "id": "decentralized-model-hosting", "name": "Mesh hosting", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "open-weights" }, { "to": "on-premise-ai" }, { "to": "ai-model-takedown" } ], "track": "today", "i18n": { "en": { "fullName": "Model mesh hosting", "factExplain": "A way to host model files across many different computers.", "humanExplain": "Mesh hosting is like a school group chat with the homework answers. Delete one message, and three kids still have screenshots.\n\nIt helps share and back up open models. Downloads can stay steady, but takedowns get messy.", "humanExplainDisplay": "Mesh hosting is like\na ==school group chat==\nwith the homework answers.\nDelete one message,\nand three kids still have ==screenshots==.\n\nIt helps share and back up open models.\nDownloads can stay steady,\nbut takedowns get messy.", "relationsNarrative": "Open-source-model\nMesh hosting helps share open-source models without one main platform.\n\nOpen weights\nOpen weights can keep spreading through many hosting nodes.\n\nOn-premise AI\nMesh hosting helps people get models for local deployment.\n\nModel takedown\nMesh hosting makes one clean model takedown much harder.", "relations": { "open-source-model": { "label": "helps spread …", "note": "Mesh hosting helps open-source models reach more places." }, "open-weights": { "label": "keeps … moving", "note": "Open weights are easier to share across many nodes." }, "on-premise-ai": { "label": "supports …", "note": "Mesh hosting makes models easier to get for local use." }, "ai-model-takedown": { "label": "makes … harder", "note": "Many hosting nodes make one clean takedown much harder." } } }, "zh": { "fullName": "Model mesh hosting / 去中心化模型托管", "factExplain": "把模型文件分散托管在多个节点上的分发方式。", "humanExplain": "不是一家医院存病历，而像街坊人手一份偏方：你收走这本，隔壁还在传抄。\n\n常用于开源模型分发和备份；获取更稳，但统一删除和治理更麻烦。", "humanExplainDisplay": "不是一家医院存病历，\n而像街坊人手一份==偏方==：\n你收走这本，\n隔壁还在==传抄==。\n\n常用于开源模型分发\n和备份；\n获取更稳，\n但统一删除和治理更麻烦。", "relationsNarrative": "Open-source-model\n它常用于分发开源模型，减少对单一平台依赖。\n\nOpen-weights\n权重公开后，更容易通过分散节点持续传播。\n\nOn-premise AI\n它让个人和企业更方便拿到模型做本地部署。\n\nAi-model-takedown\n模型一旦分散托管，统一删除和封禁会更难。", "relations": { "open-source-model": { "label": "常用于分发…", "note": "开源模型常靠它扩散到更多节点。" }, "open-weights": { "label": "承载…传播", "note": "权重一旦公开，就更适合分散托管。" }, "on-premise-ai": { "label": "支持本地部署", "note": "分散托管让本地拿模型更方便。" }, "ai-model-takedown": { "label": "增加…难度", "note": "节点分散后，统一下架更麻烦。" } } } } }, { "id": "decision-tree", "name": "Decision Tree", "layer": "L2", "era": "1963", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "cart" }, { "to": "classification" }, { "to": "regression" }, { "to": "gradient-boosting" } ], "track": "history", "i18n": { "en": { "fullName": "Decision Tree", "factExplain": "A tree-shaped model asks yes-or-no questions step by step to predict an answer.", "humanExplain": "A decision tree is like a school cafeteria line. Pizza or tacos? Milk or juice? Small questions lead to one tray.\n\nIt can sort things into groups or predict numbers. It is easy to explain, but too many splits can make it memorize the homework.", "humanExplainDisplay": "A decision tree is like a ==school cafeteria line==.\nPizza or tacos?\nMilk or juice?\n==Small questions== lead to one tray.\n\nIt can sort things into groups\nor predict numbers.\nIt is easy to explain,\nbut too many splits can make it\nmemorize the homework.", "relationsNarrative": "CART\nCART is the most common and classic way to build a Decision Tree.\n\nClassification\nA Decision Tree can do Classification by sending items into groups.\n\nRegression\nA Decision Tree can do Regression by predicting numbers, like house prices.\n\nGradient Boosting\nGradient Boosting often links many small Decision Trees to get better results.", "relations": { "cart": { "label": "is often built as …", "note": "CART is the most common kind of Decision Tree." }, "classification": { "label": "can do …", "note": "A Decision Tree can put items into different groups." }, "regression": { "label": "can also do …", "note": "A Decision Tree can predict numbers, like a house price." }, "gradient-boosting": { "label": "is a base learner for …", "note": "Many boosting methods stack small trees to improve results." } } }, "zh": { "fullName": "决策树", "factExplain": "一种按条件逐层分裂来做预测的树状模型。", "humanExplain": "决策树像网购挑手机：先看预算，再看拍照，接着看续航，一路问下去，最后落到一个选项。\n\n它常做分类和回归，好懂也好解释，但分太细容易过拟合。", "humanExplainDisplay": "决策树像网购挑手机：\n先看==预算==，再看拍照，\n接着看续航，\n一路问下去，最后落到==一个选项==。\n\n它常做分类和回归，\n好懂也好解释，\n但分太细容易过拟合。", "relationsNarrative": "CART\nCART 是决策树最常见、最经典的实现方式。\n\nClassification\n决策树常用于分类任务，按条件把样本分到类别里。\n\nRegression\n决策树也能做回归，预测房价这类连续数值。\n\nGradient Boosting\n梯度提升常把一棵棵小决策树串起来增强效果。", "relations": { "cart": { "label": "常见实现是…", "note": "CART 是决策树里最常用的一支。" }, "classification": { "label": "可用于…任务", "note": "它能把样本分到不同类别里。" }, "regression": { "label": "也可用于…任务", "note": "它也能预测连续数值结果。" }, "gradient-boosting": { "label": "作为…的基 learner", "note": "很多提升方法都拿小树来叠加。" } } } } }, { "id": "decoder-only-transformer", "name": "Decoder-only Transformer", "layer": "L3", "era": "2018", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "autoregressive-model" }, { "to": "gpt" }, { "to": "bert" } ], "track": "history", "i18n": { "en": { "fullName": "Decoder-only Transformer", "factExplain": "A Transformer design that predicts the next token, one step at a time.", "humanExplain": "A decoder-only Transformer is like driving with only a rearview mirror. You see the road behind, but the next turn stays hidden until you get there.\n\nYou meet it in chatbots and code helpers. It is the main body plan behind many big language models.", "humanExplainDisplay": "A decoder-only Transformer is like\ndriving with only ==a rearview mirror==.\nYou see the road behind,\nbut ==the next turn stays hidden== until you get there.\n\nYou meet it in chatbots and code helpers.\nIt is the main body plan\nbehind many big language models.", "relationsNarrative": "Transformer\nA decoder-only Transformer is a core Transformer branch for making new text.\n\nAutoregressive Model\nIt predicts the next token in order, so it can keep writing.\n\nGPT\nGPT made this design the main path for large language models.\n\nBERT\nBERT reads both sides of text, so it is more for reading tasks than for writing onward.", "relations": { "transformer": { "label": "is a type of …", "note": "It is the main Transformer style for making new text." }, "autoregressive-model": { "label": "uses … style", "note": "It predicts the next token one at a time." }, "gpt": { "label": "is the design behind …", "note": "GPT is the best-known example of this design." }, "bert": { "label": "contrasts with …", "note": "BERT reads both sides of text, so it is not built for long writing." } } }, "zh": { "fullName": "仅解码器 Transformer", "factExplain": "一种按顺序预测下一个 token 的 Transformer 架构。", "humanExplain": "仅解码器像开车只看后视镜和眼前路：走过的都能回头望，但前面转弯处是什么，得等到了才知道。\n\n常用来做聊天、写作和代码生成，是大模型主力骨架。", "humanExplainDisplay": "仅解码器像开车==只看后视镜==和眼前路：\n走过的都能回头望，\n但前面转弯处是什么，\n得等==到了才知道==。\n\n常用来做聊天、\n写作和代码生成，\n是大模型主力骨架。", "relationsNarrative": "Transformer\n它是 Transformer 在生成任务上的一条核心分支。\n\nAutoregressive Model\n它按顺序预测下一个 token，天然适合连续生成。\n\nGPT\nGPT 系列把这种结构做成了大语言模型主流路线。\n\nBERT\nBERT 双向读上下文更像做理解题，不是连续写下去。", "relations": { "transformer": { "label": "属于…变体", "note": "它是 Transformer 的主流生成式分支。" }, "autoregressive-model": { "label": "采用…方式", "note": "它按顺序一个个预测下一个 token。" }, "gpt": { "label": "是…典型架构", "note": "GPT 系列就是这类结构的代表。" }, "bert": { "label": "与…相对", "note": "BERT 用双向看上下文，不擅长连续生成。" } } } } }, { "id": "deep-blue", "name": "Deep Blue", "layer": "L4", "era": "1997", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "minimax-search" }, { "to": "alpha-beta-pruning" }, { "to": "game-ai" }, { "to": "alphago" } ], "track": "history", "i18n": { "en": { "fullName": "IBM Deep Blue Chess Computer", "factExplain": "IBM’s special chess computer built to play world-class chess.", "humanExplain": "Deep Blue was a calculator wearing a chess-club hoodie. It just counted moves until the room got very quiet.\n\nIt showed search plus scoring could beat top chess play. It helped push Game AI into the spotlight.", "humanExplainDisplay": "Deep Blue was a ==calculator==\nwearing a ==chess-club hoodie==.\nIt just counted moves\nuntil the room got very quiet.\n\nIt showed search plus scoring\ncould beat top chess play.\nIt helped push Game AI\ninto the spotlight.", "relationsNarrative": "Minimax Search\nDeep Blue used Minimax Search to pick the safest chess move.\n\nΑ-β Pruning\nΑ-β Pruning helped Deep Blue skip branches it did not need.\n\nGame AI\nDeep Blue made Game AI a public story, not just a lab topic.\n\nAlphaGo\nDeep Blue and AlphaGo are famous moments when machines beat human champions.", "relations": { "minimax-search": { "label": "chooses moves with …", "note": "Minimax Search helped Deep Blue compare attack and defense." }, "alpha-beta-pruning": { "label": "cuts branches with …", "note": "Α-β Pruning cuts chess lines not worth checking." }, "game-ai": { "label": "became a milestone in …", "note": "Deep Blue made game-playing AI a public story." }, "alphago": { "label": "is compared with …", "note": "Both are famous moments when machines beat human champions." } } }, "zh": { "fullName": "IBM 深蓝国际象棋计算机", "factExplain": "IBM 开发的国际象棋对弈专用系统。", "humanExplain": "深蓝像巷口棋摊的冷面算盘：不讲棋品，只把亿万变化算到你手抖。\n\n它证明搜索加评估能赢顶级棋局，并推动游戏 AI。", "humanExplainDisplay": "深蓝像==巷口棋摊的冷面算盘==：\n不讲棋品，\n只把亿万变化\n算到你==手抖==。\n\n它证明搜索加评估能赢顶级棋局，\n并推动游戏 AI。", "relationsNarrative": "Minimax Search\n深蓝用它在棋局树里挑最稳的走法。\n\nΑ-β Pruning\n它帮深蓝剪掉明显不用算的分支。\n\nGame AI\n深蓝是游戏 AI 从实验室走向大众的名场面。\n\nAlphaGo\n两者都是人类冠军被机器击败的标志时刻。", "relations": { "minimax-search": { "label": "用…选棋步", "note": "用极小化极大推演攻防得失。" }, "alpha-beta-pruning": { "label": "用…剪枝", "note": "剪掉不必细算的棋局分支。" }, "game-ai": { "label": "成为…里程碑", "note": "它让游戏对弈成为大众话题。" }, "alphago": { "label": "对照后来的…", "note": "同是人机对弈的标志事件。" } } } } }, { "id": "deep-learning", "name": "Deep Learning", "layer": "L1", "era": "2006", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "alexnet" }, { "to": "transformer" }, { "to": "scaling-law" } ], "track": "history", "i18n": { "en": { "fullName": "Deep Learning", "factExplain": "A way for layered neural networks to learn patterns from data.", "humanExplain": "Deep learning is like driveway basketball. No playbook. Just thousands of shots and a very tired hoop.\n\nIt helps AI spot pictures. It helps AI understand speech. It also powers big models.", "humanExplainDisplay": "Deep learning is like ==driveway basketball==.\nNo playbook.\nJust ==thousands of shots==\nand a very tired hoop.\n\nIt helps AI spot pictures.\nIt helps AI understand speech.\nIt also powers big models.", "relationsNarrative": "Neural-network\nDeep learning is neural networks made deeper and larger.\n\nAlexNet\nAlexNet showed deep learning could beat older methods in image tasks.\n\nTransformer\nTransformer is one of the most important designs in deep learning.\n\nScaling-law\nDeep learning models often get stronger with more data and compute.", "relations": { "neural-network": { "label": "builds on …", "note": "Deep learning is basically neural networks made deeper." }, "alexnet": { "label": "got famous through …", "note": "AlexNet made deep learning a star in image tasks." }, "transformer": { "label": "grew into …", "note": "Transformer is a key deep learning architecture." }, "scaling-law": { "label": "is driven by …", "note": "Deep learning often gets stronger with more data and compute." } } }, "zh": { "fullName": "深度学习", "factExplain": "用多层神经网络从数据中学习表示的方法。", "humanExplain": "深度学习像练武不背秘籍招式，直接上擂台打成千上万场，慢慢把眼力、手感和路数都打出来。\n\n它支撑识图、语音识别和大模型，是近十年 AI 爆发的核心引擎。", "humanExplainDisplay": "深度学习像练武\n不背秘籍招式，\n直接上擂台打成千上万场，\n慢慢把==眼力、手感和路数==\n都==打出来==。\n\n它支撑识图、\n语音识别和大模型，\n是近十年 AI 爆发的核心引擎。", "relationsNarrative": "Neural-network\n深度学习本质上是把神经网络做深、做大。\n\nAlexnet\nAlexNet 证明它能在图像任务里大幅领先传统方法。\n\nTransformer\nTransformer 是深度学习时代最关键的架构之一。\n\nScaling-law\n深度学习模型常随数据和算力扩大而变强。", "relations": { "neural-network": { "label": "建立在…之上", "note": "它本质上是更深层的神经网络方法。" }, "alexnet": { "label": "被…带火", "note": "AlexNet 让它在视觉任务中一战成名。" }, "transformer": { "label": "演化出…架构", "note": "Transformer 是深度学习的重要代表架构。" }, "scaling-law": { "label": "受…驱动", "note": "做大模型时常按规模规律继续堆。" } } } } }, { "id": "deep-q-network", "name": "Deep Q-Network", "layer": "L3", "era": "2013", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "q-learning" }, { "to": "deep-reinforcement-learning" }, { "to": "bellman-equation" }, { "to": "temporal-difference-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Deep Q-Network", "factExplain": "A reinforcement learning method with a neural network for scoring actions.", "humanExplain": "DQN is like a tiny coach inside a game controller. It gives each button a secret score, then yells, “Stop hugging the wall.”\n\nIt learns those scores by trying moves and seeing what happens. You meet it in games and control tasks.", "humanExplainDisplay": "DQN is like a ==tiny coach==\ninside a game controller.\nIt gives each button\n==a secret score==,\nthen yells,\n“Stop hugging the wall.”\n\nIt learns those scores\nby trying moves\nand seeing what happens.\nYou meet it in games\nand control tasks.", "relationsNarrative": "Q-Learning\nDQN is Q-Learning with a neural network instead of a table.\n\nDeep RL\nDQN is a classic early Deep RL method.\n\nBellman Equation\nDQN uses the Bellman Equation to build its value update target.\n\nTD Learning\nDQN uses TD Learning, so it updates scores while it plays.", "relations": { "q-learning": { "label": "extends … with a neural network", "note": "DQN replaces the Q-Learning table with a neural network." }, "deep-reinforcement-learning": { "label": "is a classic … method", "note": "DQN was an early famous result in Deep RL." }, "bellman-equation": { "label": "updates targets with …", "note": "The Bellman Equation gives DQN the target score to learn." }, "temporal-difference-learning": { "label": "learns through …", "note": "DQN changes its score after comparing its guess with what happens next." } } }, "zh": { "fullName": "深度 Q 网络", "factExplain": "用神经网络近似动作价值函数的强化学习方法。", "humanExplain": "它像下棋时心里先给每一步偷偷标分：这手能不能翻盘、会不会送子，久了就更会落子。\n\n常用于游戏和控制任务，把试错经验学成动作分值。", "humanExplainDisplay": "它像下棋时心里\n先给每一步==偷偷标分==：\n这手能不能翻盘、\n会不会==送子==，\n久了就更会落子。\n\n常用于游戏和控制任务，\n把试错经验\n学成动作分值。", "relationsNarrative": "Q-Learning\n它是 Q 学习的深度版，用神经网络代替查表。\n\nDeep-Reinforcement-Learning\n它是深度强化学习的经典代表方法之一。\n\nBellman Equation\n它用贝尔曼方程来构造价值更新目标。\n\nTD Learning\n它属于时序差分学习路线，边走边改估值。", "relations": { "q-learning": { "label": "用神经网络扩展…", "note": "它把表格版 Q 学习换成了函数逼近。" }, "deep-reinforcement-learning": { "label": "属于…代表方法", "note": "它是深度强化学习早期标志性成果。" }, "bellman-equation": { "label": "用…更新目标", "note": "它的价值更新核心来自贝尔曼方程。" }, "temporal-difference-learning": { "label": "属于…一类方法", "note": "它靠当前估计和下一步回报做差学习。" } } } } }, { "id": "deep-reinforcement-learning", "name": "Deep RL", "layer": "L2", "era": "2013", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "deep-learning" }, { "to": "deep-q-network" }, { "to": "proximal-policy-optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Deep Reinforcement Learning", "factExplain": "Reinforcement learning done with deep neural networks.", "humanExplain": "Deep RL is like training a video-game dog with treats. It tries a jump, bonks a wall, then learns the level.\n\nYou meet it in games and robots. Deep nets help it choose actions after lots of trial and error.", "humanExplainDisplay": "Deep RL is like training a ==video-game dog== with treats.\nIt tries a jump,\n==bonks a wall==,\nthen learns the level.\n\nYou meet it in games and robots.\nDeep nets help it choose actions\nafter lots of trial and error.", "relationsNarrative": "RL\nDeep RL adds deep neural networks to the RL setup.\n\nDeep Learning\nDeep Learning lets it handle images and other complex input.\n\nDQN\nDQN was an early classic method in Deep RL.\n\nppo\nppo is a common way to train these systems.", "relations": { "reinforcement-learning": { "label": "builds on …", "note": "Deep RL adds deep neural networks to RL." }, "deep-learning": { "label": "uses … for input", "note": "Deep Learning helps it handle images and other complex input." }, "deep-q-network": { "label": "classic example is …", "note": "DQN was an early famous Deep RL method." }, "proximal-policy-optimization": { "label": "often trains with …", "note": "PPO is a common modern training method for these systems." } } }, "zh": { "fullName": "Deep Reinforcement Learning（深度强化学习）", "factExplain": "用深度神经网络实现强化学习的方法。", "humanExplain": "它像新人跑业务：这单成了就记套路，碰壁了就换话术，客户见多了自然会谈。\n\n常用于游戏、机器人等决策任务，靠试错学会选动作。", "humanExplainDisplay": "它像新人跑==业务==：\n这单成了就记套路，\n碰壁了就换==话术==，\n客户见多了自然会谈。\n\n常用于游戏、机器人等决策任务，\n靠试错学会选动作。", "relationsNarrative": "Reinforcement-learning\n它是在强化学习框架里引入深度神经网络。\n\nDeep-learning\n深度学习让它能处理图像等复杂高维输入。\n\nDqn\nDQN 是深度强化学习的经典早期代表方法。\n\nPpo\nPPO 是训练这类系统时常见的优化方法。", "relations": { "reinforcement-learning": { "label": "建立在…之上", "note": "它把强化学习和深度网络结合起来。" }, "deep-learning": { "label": "借助…做表示", "note": "深度网络帮它处理高维复杂输入。" }, "deep-q-network": { "label": "经典代表是…", "note": "DQN 是它早期出圈的代表方法。" }, "proximal-policy-optimization": { "label": "常用…来训练", "note": "PPO 是现代常见训练算法之一。" } } } } }, { "id": "deepfake", "name": "Deepfake", "layer": "L6", "era": "2017", "publishedAt": "2026-05-23T11:10:00Z", "relations": [ { "to": "diffusion" }, { "to": "multimodal" }, { "to": "copyright" }, { "to": "ai-regulation" } ], "i18n": { "en": { "fullName": "Deepfake", "factExplain": "AI-made or AI-edited fake faces, voices, and videos that seem real.", "humanExplain": "A deepfake is a costume shop for pixels. It can put a movie star’s face on Uncle Bob.\n\nIt can help movies and art. It can also impersonate people for scams or fake outrage.", "humanExplainDisplay": "A deepfake is a ==costume shop for pixels==.\nIt can put a ==movie star’s face== on Uncle Bob.\n\nIt can help movies and art.\nIt can also impersonate people\nfor scams or fake outrage.", "relationsNarrative": "Diffusion\nDiffusion makes deepfake images look more real.\n\nMultimodal AI\nMultimodal lets deepfakes spread to voice, video, and identity fakes.\n\nCopyright\nDeepfakes can cause Copyright fights over faces and content ownership.\n\nAI-regulation\nAI-regulation sets limits on high-risk deepfake use.", "relations": { "diffusion": { "label": "relies on …", "note": "Diffusion can make deepfake images look more real." }, "multimodal": { "label": "uses … to fake more", "note": "Multimodal AI helps deepfakes copy faces, voices, and video." }, "copyright": { "label": "runs into …", "note": "Deepfakes can start fights over Copyright, faces, and ownership." }, "ai-regulation": { "label": "is a focus of …", "note": "AI-regulation puts limits on risky deepfake use." } } }, "zh": { "fullName": "深度伪造", "factExplain": "用 AI 生成或篡改逼真人脸、声音和视频的技术。", "humanExplain": "深度伪造像给骗子开了美颜和变声器，视频里点头的“领导”也可能是替身。\n\n它常见于换脸、仿声和假新闻，重点风险是冒充、诈骗与舆论操控。", "humanExplainDisplay": "深度伪造像==给骗子开了美颜和变声器==，\n视频里点头的“领导”\n也可能是==替身==。\n\n它常见于换脸、仿声和假新闻，\n重点风险是冒充、诈骗与舆论操控。", "relationsNarrative": "Diffusion\nDiffusion 提升了 Deepfake 在图像生成上的真实感。\n\nMultimodal AI\nMultimodal 让 Deepfake 扩展到声音、视频和身份伪造。\n\nCopyright\nDeepfake 容易触发 Copyright、肖像权和内容归属问题。\n\nAI-regulation\nAI-regulation 针对 Deepfake 的高风险使用设定限制。", "relations": { "diffusion": { "label": "依赖…" }, "multimodal": { "label": "依赖…生成" }, "copyright": { "label": "触及…" }, "ai-regulation": { "label": "是…的重点" } } } } }, { "id": "deepseek", "name": "DeepSeek", "layer": "L3", "era": "2023", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "reasoning-model" }, { "to": "open-weights" }, { "to": "qwen" } ], "track": "history", "i18n": { "en": { "fullName": "DeepSeek model / assistant", "factExplain": "A Chinese AI brand that makes large language models and an AI assistant.", "humanExplain": "DeepSeek is like the new kid at the science fair. It shows up with a budget laptop and makes the fancy booths sweat.\n\nPeople use it to chat and write code. It also made Chinese AI models harder to ignore.", "humanExplainDisplay": "DeepSeek is like the ==new kid at the science fair==.\nIt shows up with a budget laptop\nand ==makes the fancy booths sweat==.\n\nPeople use it to chat and write code.\nIt also made Chinese AI models harder to ignore.", "relationsNarrative": "LLM\nDeepSeek is an LLM. People use it to chat and write code.\n\nReasoning-model\nSome DeepSeek versions drew attention for strong reasoning.\n\nOpen weights\nDeepSeek released some weights and pushed talk about open AI costs.\n\nQwen\nDeepSeek and Qwen are rivals among Chinese AI models.", "relations": { "llm": { "label": "is a kind of …", "note": "DeepSeek is basically a large language model you can talk to." }, "reasoning-model": { "label": "is often seen as …", "note": "Some DeepSeek versions got famous for strong reasoning." }, "open-weights": { "label": "sparked talk about …", "note": "Its released weights fueled talk about open AI and cost." }, "qwen": { "label": "competes with …", "note": "DeepSeek and Qwen are both big Chinese AI model names." } } }, "zh": { "fullName": "深度求索模型 / 助手", "factExplain": "一家推出大模型和 AI 助手的中国 AI 品牌。", "humanExplain": "像武林里突然冒头的少侠：出手利索，招式还硬，几场切磋打下来，连江湖上的报价规矩都被它搅动了。\n\n常被用来聊天和写代码，也让国产大模型更受关注。", "humanExplainDisplay": "像武林里突然冒头的==少侠==：\n出手利索，\n招式还硬，\n几场切磋打下来，连==报价规矩都被它搅动了==。\n\n常被用来聊天和写代码，\n也让国产大模型更受关注。", "relationsNarrative": "LLM\n它本质上属于大语言模型，可用于对话、写作和编码。\n\nReasoning-model\n它的部分版本因推理表现受到关注，常被归入推理模型讨论。\n\nOpen-weights\n它公开部分模型权重，推动了开放生态与成本讨论。\n\nQwen\n它与 Qwen 同属国产大模型阵营，常被用户并列比较。", "relations": { "llm": { "label": "属于…一类", "note": "它本质上是可对话的大语言模型。" }, "reasoning-model": { "label": "常被视作…代表", "note": "其部分版本因推理能力出圈。" }, "open-weights": { "label": "推动…讨论", "note": "公开权重策略拉高行业关注度。" }, "qwen": { "label": "与…同场竞争", "note": "两者都是国产大模型重要玩家。" } } } } }, { "id": "dendral", "name": "DENDRAL", "layer": "L4", "era": "1965", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "expert-system" }, { "to": "symbolic-ai" }, { "to": "knowledge-representation" }, { "to": "mycin" } ], "track": "history", "i18n": { "en": { "fullName": "DENDRAL", "factExplain": "An early expert system that used rules to infer molecule structures.", "humanExplain": "DENDRAL was a crime lab for molecules. It read a molecule’s lab fingerprint and drew the suspect.\n\nChemists used it to work out molecule structures. It helped start expert systems and rule-based knowledge in AI.", "humanExplainDisplay": "DENDRAL was a ==crime lab== for molecules.\nIt read a molecule’s ==lab fingerprint==\nand drew the suspect.\n\nChemists used it to work out molecule structures.\nIt helped start expert systems\nand rule-based knowledge in AI.", "relationsNarrative": "Expert System\nDENDRAL was an early landmark case for expert systems.\n\nSymbolic AI\nDENDRAL used clear rules and search, not neural networks.\n\nKR\nChemistry knowledge was written as rules for the program to reason with.\n\nMYCIN\nMYCIN carried on the same expert-system style.", "relations": { "expert-system": { "label": "helped launch …", "note": "DENDRAL was one of the first successful expert systems." }, "symbolic-ai": { "label": "used …", "note": "It used clear rules and search, not neural nets." }, "knowledge-representation": { "label": "relied on …", "note": "Chemistry know-how was written as rules the program could use." }, "mycin": { "label": "influenced …", "note": "MYCIN followed Stanford’s expert-system path." } } }, "zh": { "fullName": "有机化学结构解析专家系统", "factExplain": "早期用规则推断分子结构的专家系统。", "humanExplain": "DENDRAL 是化学版福尔摩斯：盯着质谱线索，倒推出分子长相。\n\n用于化学结构解析，开创专家系统和知识表示实践。", "humanExplainDisplay": "DENDRAL 是\n==化学版福尔摩斯==：\n盯着质谱线索，\n倒推出分子长相。\n\n用于化学结构解析，\n开创专家系统，\n和知识表示实践。", "relationsNarrative": "Expert System\nDENDRAL 是专家系统的早期标志性案例。\n\nSymbolic AI\n它靠显式规则和搜索，而不是神经网络。\n\nKnowledge Representation\n化学知识被写成规则，供程序推理。\n\nMYCIN\nMYCIN 继承了这类专家系统思路。", "relations": { "expert-system": { "label": "开创…范式", "note": "它是最早成功的专家系统之一。" }, "symbolic-ai": { "label": "采用…路线", "note": "用规则和搜索模拟专家判断。" }, "knowledge-representation": { "label": "依赖…", "note": "化学经验被编码成可推理规则。" }, "mycin": { "label": "影响…", "note": "MYCIN 延续了斯坦福专家系统路线。" } } } } }, { "id": "denoising-autoencoder", "name": "Denoising Autoencoder", "layer": "L3", "era": "2008", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "autoencoder" }, { "to": "representation-learning" }, { "to": "self-supervised-learning" }, { "to": "regularization" } ], "track": "history", "i18n": { "en": { "fullName": "Denoising Autoencoder", "factExplain": "An Autoencoder that learns by fixing a noisy version of its input.", "humanExplain": "A Denoising Autoencoder is like cleaning a ketchup-smudged school photo. First we add the ketchup. Then it learns to find the face anyway.\n\nYou meet it in pretraining, cleanup, and weird-data spotting. It learns features that still work when the input gets messy.", "humanExplainDisplay": "A Denoising Autoencoder is like cleaning\na ==ketchup-smudged school photo==.\nFirst we ==add the ketchup==.\nThen it learns to find the face anyway.\n\nYou meet it in pretraining,\ncleanup,\nand weird-data spotting.\nIt learns features that still work\nwhen the input gets messy.", "relationsNarrative": "Autoencoder\nIt adds noise to an Autoencoder input, then learns to rebuild it.\n\nRepresentation Learning\nIt learns stable features instead of copying the input.\n\nSSL\nIt makes fake noise, then uses the clean input as the answer.\n\nRegularization\nNoise training works like Regularization and reduces memorizing.", "relations": { "autoencoder": { "label": "adds noise to …", "note": "It is the noise-ready version of an Autoencoder." }, "representation-learning": { "label": "learns steadier …", "note": "Cleaning noisy input forces it to keep the key features." }, "self-supervised-learning": { "label": "uses an … task", "note": "The fake noise gives the model its own training target." }, "regularization": { "label": "acts like …", "note": "Noise during training helps it stop memorizing." } } }, "zh": { "fullName": "降噪自编码器", "factExplain": "一种通过还原加噪输入学习表示的自编码器。", "humanExplain": "降噪自编码器像老照片修复：先故意撒灰，再练会把人脸擦亮。\n\n用于预训练、去噪和异常检测，学到更抗噪的特征。", "humanExplainDisplay": "降噪自编码器像老照片修复：\n先==故意撒灰==，\n再练会把人脸==擦亮==。\n\n用于预训练、去噪和异常检测，\n学到更抗噪的特征。", "relationsNarrative": "Autoencoder\n它是在自编码器输入上加噪，再学习还原。\n\nRepresentation Learning\n它逼模型抓住稳定特征，而不是照抄输入。\n\nSelf-Supervised Learning\n造噪声再还原，让数据自己提供训练信号。\n\nRegularization\n加噪训练像一种正则化，减少死记硬背。", "relations": { "autoencoder": { "label": "在…上加噪训练", "note": "它是自编码器的抗噪版本。" }, "representation-learning": { "label": "学习更稳的…", "note": "还原噪声输入，逼模型抓关键特征。" }, "self-supervised-learning": { "label": "用…式任务训练", "note": "造噪声再还原，本身就是监督信号。" }, "regularization": { "label": "起到…效果", "note": "加噪训练能减少死记硬背。" } } } } }, { "id": "denoising-diffusion-probabilistic-model", "name": "DDPM", "layer": "L3", "era": "2020", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "generative-model" }, { "to": "u-net" }, { "to": "classifier-free-guidance" } ], "track": "history", "i18n": { "en": { "fullName": "Denoising Diffusion Probabilistic Model", "factExplain": "A diffusion model that makes data by removing noise step by step.", "humanExplain": "DDPM starts with a screen full of TV static. Then it cleans the mess, step by step, until a puppy appears.\n\nIt often powers text-to-image tools and photo edits. Results are steady, but it needs many steps.", "humanExplainDisplay": "DDPM starts with a screen full of ==TV static==.\nThen it cleans the mess,\nstep by step,\nuntil ==a puppy appears==.\n\nIt often powers text-to-image tools\nand photo edits.\nResults are steady,\nbut it needs many steps.", "relationsNarrative": "Diffusion\nDDPM is a classic Diffusion model.\n\nGenerative Model\nDDPM learns data patterns and makes new samples.\n\nU-Net\nU-Net often acts as the denoising network in DDPM.\n\nCFG\nCFG often guides DDPM toward the prompt.", "relations": { "diffusion": { "label": "sets classic form for …", "note": "DDPM is a classic model in the diffusion family." }, "generative-model": { "label": "belongs to …", "note": "It learns data patterns and makes new samples." }, "u-net": { "label": "denoises with …", "note": "U-Net often predicts the noise to remove." }, "classifier-free-guidance": { "label": "is guided by …", "note": "CFG helps DDPM follow the prompt more closely." } } }, "zh": { "fullName": "Denoising Diffusion Probabilistic Model，去噪扩散概率模型", "factExplain": "通过逐步去噪生成数据的扩散模型。", "humanExplain": "DDPM像在窗上哈满白雾：先糊住世界，再一层层擦出图像。\n\n常支撑文生图和修图；画质稳，但生成步数多。", "humanExplainDisplay": "DDPM像在窗上\n==哈满白雾==：\n先糊住世界，\n再一层层==擦出图像==。\n\n常支撑文生图和修图；\n画质稳，\n但生成步数多。", "relationsNarrative": "Diffusion\nDDPM 是扩散模型的经典概率化实现。\n\nGenerative Model\nDDPM 学习数据分布，用来生成新样本。\n\nU-Net\nU-Net 常作为 DDPM 的去噪网络骨架。\n\nClassifier-free Guidance\n它常用于引导 DDPM 更贴近提示词。", "relations": { "diffusion": { "label": "奠定…经典形式", "note": "DDPM 是扩散路线的代表模型。" }, "generative-model": { "label": "属于…", "note": "它通过学分布来生成新样本。" }, "u-net": { "label": "常用…去噪", "note": "U-Net 常负责预测噪声。" }, "classifier-free-guidance": { "label": "用…引导生成", "note": "引导让结果更贴近提示词。" } } } } }, { "id": "deployment-simulation", "name": "Deployment Simulation", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "agent-harness" }, { "to": "agent-security" }, { "to": "llmops" }, { "to": "computer-use" } ], "track": "today", "i18n": { "en": { "fullName": "Deployment Simulation", "factExplain": "A pre-launch test in a practice copy of the AI system’s real setup.", "humanExplain": "Deployment Simulation is a dress rehearsal for your AI. Let it trip over fake scenery, not the real checkout button.\n\nTeams use it before launch to stress-test Agents. They also test tool calls and permission limits. It finds the crash points early.", "humanExplainDisplay": "Deployment Simulation is a ==dress rehearsal== for your AI.\nLet it trip over ==fake scenery==,\nnot the real checkout button.\n\nTeams use it before launch\nto stress-test Agents.\nThey also test tool calls\nand permission limits.\nIt finds the crash points early.", "relationsNarrative": "Agent harness\nDeployment Simulation often uses an Agent harness to replay many real tasks.\n\nAgent Security\nDeployment Simulation can expose risky permissions and bad actions early.\n\nLLMOps\nDeployment Simulation is a pre-launch check inside LLMOps.\n\nComputer use\nComputer use tasks need practice runs in a real-like setup.", "relations": { "agent-harness": { "label": "often runs through …", "note": "An Agent harness can run many practice tasks at once." }, "agent-security": { "label": "helps … spot risks", "note": "It can reveal risky permissions before real users arrive." }, "llmops": { "label": "fits into …", "note": "It is a key pre-launch check for AI systems." }, "computer-use": { "label": "often tests …", "note": "It is useful for web and software tasks." } } }, "zh": { "fullName": "Deployment Simulation｜部署模拟", "factExplain": "在正式上线前模拟 AI 系统真实运行环境的测试方法。", "humanExplain": "跟武侠里“闭门试招”一个道理：先拿木人桩过一遍，别真出门就把人、钱、流程全打乱。\n\n常用于上线前压测 Agent、工具调用和权限边界，提前找翻车点。", "humanExplainDisplay": "跟武侠里\n==闭门试招==一个道理：\n先拿木人桩过一遍，\n别真出门就把==人、钱、流程全打乱==。\n\n常用于上线前压测 Agent、\n工具调用和权限边界，\n提前找翻车点。", "relationsNarrative": "Agent Harness\n部署模拟常接入测试框架，批量复现真实任务。\n\nAgent Security\n它能提前暴露越权、误操作等安全风险。\n\nLLMOps\n它属于 AI 系统上线前的重要验证环节。\n\nComputer Use\n电脑操作类任务最需要先做真实环境模拟。", "relations": { "agent-harness": { "label": "常接入…测试", "note": "常靠测试框架批量跑模拟任务。" }, "agent-security": { "label": "帮…提前排雷", "note": "可提前发现权限滥用等风险。" }, "llmops": { "label": "属于…环节", "note": "是上线前验证的重要步骤。" }, "computer-use": { "label": "常模拟…场景", "note": "尤其适合网页和软件操作任务。" } } } } }, { "id": "description-logic", "name": "DL", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "ontology" }, { "to": "knowledge-graph" } ], "track": "history", "i18n": { "en": { "fullName": "Description Logic", "factExplain": "A logic system for writing concepts, links, and rules in a formal way.", "humanExplain": "Description Logic is a rulebook for a picky school club. It says who is in before anyone prints the T-shirts.\n\nIt helps computers model knowledge and catch rules that clash. You meet it in ontologies, knowledge graphs, and smart search.", "humanExplainDisplay": "Description Logic is a ==rulebook==\nfor a picky school club.\nIt says ==who is in==\nbefore anyone prints the T-shirts.\n\nIt helps computers model knowledge\nand catch rules that clash.\nYou meet it in ontologies,\nknowledge graphs,\nand smart search.", "relationsNarrative": "KR\nDescription Logic is an important formal method in KR.\n\nOntology\nMany ontology languages build on Description Logic.\n\nKnowledge Graph\nDescription Logic adds type, link, and constraint rules to a Knowledge Graph.", "relations": { "knowledge-representation": { "label": "is a … method", "note": "Description Logic is a classic formal tool in KR." }, "ontology": { "label": "supports …", "note": "Many ontology languages use Description Logic as their logic base." }, "knowledge-graph": { "label": "adds rules to …", "note": "It lets a graph check types, links, and rule conflicts." } } }, "zh": { "fullName": "Description Logic／描述逻辑", "factExplain": "一种用于形式化表示概念、关系与约束的逻辑体系。", "humanExplain": "描述逻辑像给武林门派立规矩：谁算同门、能练哪路功、犯了哪条算冲突，都得写死。\n\n常用于知识建模和一致性检查；适合图谱、语义检索等规则场景。", "humanExplainDisplay": "描述逻辑像给武林门派\n==立规矩==：\n谁算同门、能练哪路功，\n犯了哪条算==冲突==，\n都得写死。\n\n常用于知识建模和\n一致性检查；\n适合图谱、语义检索等规则场景。", "relationsNarrative": "Knowledge Representation\n描述逻辑是知识表示领域的重要形式化方法。\n\nOntology\n很多本体语言建立在描述逻辑的表达能力之上。\n\nKnowledge Graph\n它能为知识图谱补上类型、关系和约束规则。", "relations": { "knowledge-representation": { "label": "属于…方法", "note": "它是知识表示里的经典形式化工具。" }, "ontology": { "label": "支撑…表达", "note": "很多本体语言以它为逻辑基础。" }, "knowledge-graph": { "label": "给…加规则", "note": "让图谱不只存事实，还能校验约束。" } } } } }, { "id": "deterministic-ai", "name": "Deterministic AI", "layer": "L5", "sublayer": "architecture", "era": "1950s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "temperature" }, { "to": "symbolic-ai" }, { "to": "structured-output" }, { "to": "hallucination" } ], "track": "today", "i18n": { "en": { "fullName": "Deterministic AI", "factExplain": "An AI that gives the same output for the same input and settings.", "humanExplain": "Deterministic AI is like a vending machine. Press B7, get the same sad granola bar every time.\n\nYou see it in approvals and risk checks. You also see it in factory controls. The goal is repeatability, not winging it.", "humanExplainDisplay": "Deterministic AI is like a ==vending machine==.\nPress B7,\nget the ==same sad granola bar== every time.\n\nYou see it in approvals and risk checks.\nYou also see it in factory controls.\nThe goal is repeatability,\nnot winging it.", "relationsNarrative": "Temperature\nLower Temperature usually makes AI answers more repeatable.\n\nSymbolic AI\nSymbolic AI uses rules, so it is often easier to predict.\n\nStructured output\nStructured output keeps the model inside a fixed format.\n\nHallucination\nDeterministic AI can repeat a false answer too.", "relations": { "temperature": { "label": "controls randomness with …", "note": "Low Temperature makes answers closer to repeatable." }, "symbolic-ai": { "label": "often appears in …", "note": "Rule reasoning is easier to predict." }, "structured-output": { "label": "limits results with …", "note": "A fixed format makes outputs steadier." }, "hallucination": { "label": "does not erase …", "note": "A repeatable answer can still be false." } } }, "zh": { "fullName": "确定性 AI", "factExplain": "固定条件下，同一输入产生相同输出。", "humanExplain": "确定性 AI 是象棋棋谱：同一盘同一步，照谱走永远落同一格。\n\n适合审批、风控、工控，重在可复现，不求临场发挥。", "humanExplainDisplay": "确定性 AI 是==象棋棋谱==：\n==同一盘同一步==，\n照谱走永远落同一格。\n\n适合审批、风控、工控，\n重在可复现，\n不求临场发挥。", "relationsNarrative": "Temperature\nTemperature 越低，生成结果通常越接近可复现。\n\nSymbolic AI\nSymbolic AI 依赖规则推理，天然更容易确定。\n\nStructured output\nStructured output 用固定格式约束模型输出范围。\n\nHallucination\n确定性不等于真实，错误也可能稳定复现。", "relations": { "temperature": { "label": "用…控制随机性", "note": "低温度让生成更接近可复现。" }, "symbolic-ai": { "label": "常见于…", "note": "规则推理天然更可预测。" }, "structured-output": { "label": "配合…约束结果", "note": "固定格式让输出更稳定。" }, "hallucination": { "label": "不等于消灭…", "note": "可复现不代表一定真实。" } } } } }, { "id": "dexterous-robotic-hand", "name": "Dexterous robotic hand", "layer": "L5", "sublayer": "product", "era": "1980s", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "robot-cerebellum" } ], "track": "today", "i18n": { "en": { "fullName": "Dexterous Robotic Hand", "factExplain": "A multi-finger robot hand for careful grabbing and moving objects.", "humanExplain": "A dexterous robotic hand is a robot's pancake-flipping hand. It can pinch a Lego and hold an egg without making breakfast sad.\n\nIt helps robots do tiny careful jobs. You may see it in factories or care robots. One day, it may fold your laundry.", "humanExplainDisplay": "A dexterous robotic hand is a robot's ==pancake-flipping hand==.\nIt can ==pinch a Lego==\nand hold an egg\nwithout making breakfast sad.\n\nIt helps robots do tiny careful jobs.\nYou may see it in factories or care robots.\nOne day, it may fold your laundry.", "relationsNarrative": "Robotics\nA dexterous robotic hand is the end part a robot uses to touch objects.\n\nEmbodied AI\nIt lets Embodied AI reach out and touch the world.\n\nVLA\nA VLA can turn what it sees and hears into hand commands.\n\nRobot Cerebellum\nThe Robot Cerebellum keeps finger moves steady and smooth.", "relations": { "robotics": { "label": "serves as …'s hand", "note": "It is the part a robot uses to touch things." }, "embodied-ai": { "label": "gives action to …", "note": "Embodied AI needs a body to act in the world." }, "vision-language-action-model-vla": { "label": "takes commands from …", "note": "A VLA can turn a goal into hand movements." }, "robot-cerebellum": { "label": "depends on … for control", "note": "Small finger moves need steady control." } } }, "zh": { "fullName": "灵巧机器人手", "factExplain": "一种用于精细抓取与操作的多指机器人末端执行器。", "humanExplain": "灵巧机械手有煎饼摊手艺：能夹能捏能翻面，还知道轻重火候。\n\n用于装配、护理和家务，让机器人能完成细活。", "humanExplainDisplay": "灵巧机械手有\n==煎饼摊手艺==：\n能夹能捏能翻面，\n还知道==轻重火候==。\n\n用于装配、护理和家务，\n让机器人能完成细活。", "relationsNarrative": "Robotics\n它是机器人接触物体的末端执行器。\n\nEmbodied AI\n它让具身智能真正伸手碰世界。\n\nVLA\nVLA 可把视觉和语言目标转成动作指令。\n\nRobot Cerebellum\n它负责把手指动作控制得更稳更细。", "relations": { "robotics": { "label": "作为…末端", "note": "它是机器人接触世界的手。" }, "embodied-ai": { "label": "承载…动作", "note": "具身智能要靠身体执行。" }, "vision-language-action-model-vla": { "label": "接收…指令", "note": "VLA 可把目标转成动作。" }, "robot-cerebellum": { "label": "依赖…控稳", "note": "细动作需要稳定协调。" } } } } }, { "id": "differential-privacy", "name": "Differential Privacy", "layer": "L2", "era": "2006", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "federated-learning" }, { "to": "regularization" } ], "track": "history", "i18n": { "en": { "fullName": "Differential Privacy", "factExplain": "A method that hides one person’s data inside group statistics.", "humanExplain": "Differential Privacy is like a school lunch poll with a little static added. You see pizza is winning, but not which kid voted for it.\n\nIt is used in data analysis and model training. It protects personal details, but the numbers get a bit less exact.", "humanExplainDisplay": "Differential Privacy is like a ==school lunch poll==\nwith a little ==static added==.\nYou see pizza is winning,\nbut not which kid voted for it.\n\nIt is used in data analysis\nand model training.\nIt protects personal details,\nbut the numbers get a bit less exact.", "relationsNarrative": "Data-privacy\nDifferential Privacy is a classic way to protect personal data.\n\nFL\nFL often adds Differential Privacy to lower the risk of leaks during training.\n\nRegularization\nDifferential Privacy trades a little accuracy for stronger limits and protection.", "relations": { "data-privacy": { "label": "protects … with math", "note": "It uses math to protect one person’s data." }, "federated-learning": { "label": "often pairs with …", "note": "FL often adds it to reduce data leaks." }, "regularization": { "label": "adds limits like …", "note": "Both trade a little accuracy for stronger limits." } } }, "zh": { "fullName": "Differential Privacy｜差分隐私", "factExplain": "一种在统计结果中隐藏个体信息的隐私保护方法。", "humanExplain": "差分隐私像相亲局只放年龄段和收入档，你看得出大盘行情，就是锁不准某个人。\n\n常用于数据分析和模型训练，保护个人信息，但结果通常会牺牲一点精确度。", "humanExplainDisplay": "差分隐私像相亲局只放==年龄段和收入档==，\n你看得出大盘行情，\n就是==锁不准某个人==。\n\n常用于数据分析和模型训练，\n保护个人信息，\n但结果通常会牺牲一点精确度。", "relationsNarrative": "Data-privacy\n它是实现数据隐私保护的一种经典技术路线。\n\nFL\n联邦学习常结合它，降低训练中泄露个体信息的风险。\n\nRegularization\n它也会牺牲一点精度，换取更强的约束与保护。", "relations": { "data-privacy": { "label": "是…的技术手段", "note": "它用数学方式保护个体数据。" }, "federated-learning": { "label": "常与…搭配", "note": "联邦训练常叠加它进一步防泄露。" }, "regularization": { "label": "像…一样做约束", "note": "都会换一点精度，换更强约束。" } } } } }, { "id": "diffusion-language-model", "name": "Diffusion Language Model", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "autoregressive-model" }, { "to": "language-modeling" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Diffusion Language Model", "factExplain": "A language model writes by cleaning noisy text step by step.", "humanExplain": "Normal AI writes like a typewriter, one word at a time. This one starts with word soup, then fixes the whole bowl.\n\nIt can write many parts at once. It helps rewrite text and fill gaps.", "humanExplainDisplay": "Normal AI writes like a ==typewriter==,\none word at a time.\nThis one starts with ==word soup==,\nthen fixes the whole bowl.\n\nIt can write many parts at once.\nIt helps rewrite text\nand fill gaps.", "relationsNarrative": "Diffusion\nDiffusion language models bring step-by-step cleanup from Diffusion into text.\n\nAutoregressive Model\nThey do not add one token at a time. They revise the whole text in rounds.\n\nLM\nThis is a new route in LM. The goal is still text generation.\n\nLLM\nIt can be seen as another generation style for an LLM.", "relations": { "diffusion": { "label": "brings … to text", "note": "It uses the step-by-step cleanup idea from Diffusion." }, "autoregressive-model": { "label": "contrasts with …", "note": "One cleans text in rounds. The other adds one token at a time." }, "language-modeling": { "label": "offers a new route for …", "note": "It is another way to generate language." }, "llm": { "label": "can be a … variant", "note": "It is one possible path for large language models." } } }, "zh": { "fullName": "扩散语言模型", "factExplain": "把文本生成改成逐步去噪的语言模型。", "humanExplain": "不像主播一字字往外蹦，更像老师改卷：先发一张糊答案，再一轮轮修到能交卷。\n\n适合并行生成和文本修补，常用于改写、补全。", "humanExplainDisplay": "不像主播一字字往外蹦，\n更像老师==改卷==：\n先发一张糊答案，\n再一轮轮==修到能交卷==。\n\n适合并行生成和文本修补，\n常用于改写、\n补全。", "relationsNarrative": "Diffusion\n它把扩散生成那套逐步去噪思路搬到了文本上。\n\nAutoregressive Model\n它不靠逐 token 续写，而是反复修正整段文本。\n\nLanguage-modeling\n它是语言建模的一条新路线，目标仍是生成文本。\n\nLLM\n它可被视为大语言模型的另一种生成范式。", "relations": { "diffusion": { "label": "把…搬到文本", "note": "核心思路来自扩散式生成。" }, "autoregressive-model": { "label": "对比…路线", "note": "一个逐步去噪，一个逐 token 续写。" }, "language-modeling": { "label": "属于…新解法", "note": "它是在重做语言生成这件事。" }, "llm": { "label": "可作为…变体", "note": "它也是大语言模型的一种方向。" } } } } }, { "id": "diffusion", "name": "Diffusion", "layer": "L2", "era": "2020", "publishedAt": "2026-05-23T09:05:00Z", "relations": [ { "to": "multimodal" }, { "to": "prompt" }, { "to": "neural-network" } ], "i18n": { "en": { "fullName": "Diffusion Model", "factExplain": "A generative model for making images by removing noise step by step.", "humanExplain": "Diffusion is like a Polaroid slowly appearing in TV static. First fuzz, then boom: a corgi in a tiny crown.\n\nIt turns a Prompt into visual content. You meet it in AI art, image edits, and video tools.", "humanExplainDisplay": "Diffusion is like a ==Polaroid==\nslowly appearing in ==TV static==.\nFirst fuzz,\nthen boom:\na corgi in a tiny crown.\n\nIt turns a Prompt into visual content.\nYou meet it in AI art,\nimage edits,\nand video tools.", "relationsNarrative": "Multimodal AI\nDiffusion often helps Multimodal systems create visual content.\n\nPrompt\nA Prompt gives Diffusion its goal and limits.\n\nNeural-network\nA Neural-network predicts the noise Diffusion should remove.", "relations": { "multimodal": { "label": "can power …", "note": "Diffusion often powers the visual side of Multimodal generation." }, "prompt": { "label": "follows …", "note": "A Prompt points Diffusion toward the image you want." }, "neural-network": { "label": "uses …", "note": "A Neural-network predicts the noise to remove at each step." } } }, "zh": { "fullName": "扩散模型", "factExplain": "一种通过逐步去噪生成图片等内容的生成模型。", "humanExplain": "扩散模型像把电视雪花屏倒着擦，一点点擦成能发朋友圈的图。\n\n它常用于文生图、修图和视频生成，效果惊艳但细节可能乱编。", "humanExplainDisplay": "扩散模型像把==电视雪花屏==倒着擦，\n一点点擦成==能发朋友圈的图==。\n\n它常用于文生图、修图和视频生成，\n效果惊艳但细节可能乱编。", "relationsNarrative": "Multimodal AI\nMultimodal 生成视觉内容时，常依赖 Diffusion 路线。\n\nPrompt\nPrompt 为 Diffusion 指定生成方向和条件约束。\n\nNeural-network\nNeural-network 负责 Diffusion 中的噪声预测过程。", "relations": { "multimodal": { "label": "可组成…" }, "neural-network": { "label": "基于…" } } } } }, { "id": "digital-human", "name": "Digital human", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "tts" }, { "to": "deepfake" }, { "to": "content-provenance" } ], "track": "today", "i18n": { "en": { "fullName": "Digital human", "factExplain": "A virtual person made and driven by AI.", "humanExplain": "A digital human is like a video-game cashier. It smiles, talks, and never asks for a lunch break.\n\nYou meet it in livestream shops. You also see it at help desks and online shows. It saves staff time, but a fake look breaks the spell.", "humanExplainDisplay": "A digital human is like a ==video-game cashier==.\nIt smiles, talks,\nand ==never asks for a lunch break==.\n\nYou meet it in livestream shops.\nYou also see it at help desks and online shows.\nIt saves staff time,\nbut a fake look breaks the spell.", "relationsNarrative": "Multimodal AI\nA digital human often needs text, voice, and visuals to work as one.\n\nTTS\nTTS turns written words into the voice a digital human can speak.\n\nDeepfake\nIf it looks too real, a digital human can be confused with a deepfake.\n\nContent provenance\nContent provenance helps people tell if it was made by AI.", "relations": { "multimodal": { "label": "depends on …", "note": "It often needs sound, images, and text to work together." }, "tts": { "label": "speaks with …", "note": "TTS turns text into the voice it uses." }, "deepfake": { "label": "gets confused with …", "note": "A very real-looking digital human can blur into fake person content." }, "content-provenance": { "label": "needs …", "note": "Content provenance helps show if it is human-made or AI-made." } } }, "zh": { "fullName": "数字人", "factExplain": "用 AI 生成并驱动的虚拟人物形象。", "humanExplain": "数字人像电视剧里的群演替身：长得像个人，导演一喊开机，它就能张嘴、微笑、连着演十小时。\n\n它常用于带货、客服和主持，能省人力，但不够真实就容易让人出戏。", "humanExplainDisplay": "数字人像电视剧里的\n==群演替身==：长得像个人，\n导演一喊开机，\n它就能张嘴、微笑，\n连着演==十小时==。\n\n它常用于带货、\n客服和主持，\n能省人力，但不够真实\n就容易让人出戏。", "relationsNarrative": "Multimodal AI\n数字人通常要把文本、语音和画面一起协同起来。\n\nTTS\n语音合成负责把文字变成它能说出口的声音。\n\nDeepfake\n当它过于逼真时，常会和伪造真人内容混在一起。\n\nContent provenance\n来源标记能帮助用户分清它是不是 AI 生成的。", "relations": { "multimodal": { "label": "依赖…整合", "note": "它通常要同时处理声音、图像和文本。" }, "tts": { "label": "用…开口说话", "note": "语音合成让虚拟形象真正发声。" }, "deepfake": { "label": "容易被…混淆", "note": "逼真数字形象常与伪造内容边界模糊。" }, "content-provenance": { "label": "需要…标识", "note": "来源标记有助区分真人与生成内容。" } } } } }, { "id": "digital-twin", "name": "Digital twin", "layer": "L5", "sublayer": "architecture", "era": "2002", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "physical-ai-ai" }, { "to": "deployment-simulation" }, { "to": "robotics" }, { "to": "world-model" } ], "track": "today", "i18n": { "en": { "fullName": "Digital Twin", "factExplain": "A live virtual copy of a real thing, used to watch and test it.", "humanExplain": "It is like a video game copy of a real factory. If one real machine overheats, its game twin starts sweating too.\n\nTeams use it in factories and smart cities to watch the real thing. They test fixes there before touching the real one.", "humanExplainDisplay": "It is like a ==video game copy==\nof a real factory.\nIf one real machine overheats,\nits ==game twin== starts sweating too.\n\nTeams use it in factories and smart cities\nto watch the real thing.\nThey test fixes there\nbefore touching the real one.", "relationsNarrative": "Physical AI\nA digital twin gives Physical AI a safe virtual place to try things.\n\nDeployment Simulation\nDeployment Simulation uses a digital twin to rehearse real scenes early.\n\nRobotics\nRobotics can use a digital twin to practice skills with less risk.\n\nWorld model\nA World model is more internal, while a digital twin is a live outside map.", "relations": { "physical-ai-ai": { "label": "gives … a test world", "note": "Physical AI can practice inside the virtual copy first." }, "deployment-simulation": { "label": "supports …", "note": "Teams can test the site response before launch." }, "robotics": { "label": "helps … practice", "note": "Robots can fail safely in the virtual scene." }, "world-model": { "label": "mirrors the world for …", "note": "It turns the real environment into an updating model." } } }, "zh": { "fullName": "数字孪生", "factExplain": "物理对象的实时虚拟映射，用于监控和仿真。", "humanExplain": "数字孪生是会喘气的城市沙盘：路一堵、机一热，虚拟现场马上跟着变。\n\n用于监控、调度和训练，先在虚拟现场试错。", "humanExplainDisplay": "数字孪生是\n==会喘气的城市沙盘==：\n路一堵、机一热，\n虚拟现场==跟着变==。\n\n用于监控、调度和训练，\n先在虚拟现场\n试错。", "relationsNarrative": "Physical AI\n数字孪生为实体智能提供可试错的虚拟现场。\n\nDeployment Simulation\n部署仿真可借它提前演练真实场景。\n\nRobotics\n机器人用它在虚拟副本里练技能、避风险。\n\nWorld Model\n世界模型偏内部理解，它偏外部实时映射。", "relations": { "physical-ai-ai": { "label": "给…提供仿真场", "note": "物理智能常先在虚拟副本里练手。" }, "deployment-simulation": { "label": "用于…试运行", "note": "上线前可先模拟现场反应。" }, "robotics": { "label": "帮…练技能", "note": "机器人可在虚拟场景中安全试错。" }, "world-model": { "label": "外化…环境", "note": "它把真实环境做成可更新模型。" } } } } }, { "id": "dimensionality-reduction", "name": "Dim. Reduction", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "curse-of-dimensionality" }, { "to": "principal-component-analysis" }, { "to": "embedding" }, { "to": "feature-selection" } ], "track": "history", "i18n": { "en": { "fullName": "Dimensionality Reduction", "factExplain": "A way to squeeze high-dimensional data into a smaller, useful form.", "humanExplain": "Dimensionality reduction is like packing a messy backpack into one lunchbox. You keep the homework, not the mystery crumbs.\n\nIt helps draw messy data as a simple picture. It can cut noise and make models run faster.", "humanExplainDisplay": "Dimensionality reduction is like\npacking a ==messy backpack== into\none ==lunchbox==.\nYou keep the homework,\nnot the mystery crumbs.\n\nIt helps draw messy data\nas a simple picture.\nIt can cut noise\nand make models run faster.", "relationsNarrative": "Dim Curse\nDimensionality reduction helps with sparse data and heavy computing in high dimensions.\n\nPCA\nPCA is one of the most classic ways to reduce dimensions.\n\nEmbedding\nMany embedding methods learn a more compact form of data.\n\nFeature Selection\nDimensionality reduction mixes features into new ones. It does not just delete old ones.", "relations": { "curse-of-dimensionality": { "label": "helps with …", "note": "When dimensions get too high, reduction is a common fix." }, "principal-component-analysis": { "label": "often uses …", "note": "PCA is the classic linear method for dimensionality reduction." }, "embedding": { "label": "creates compact …", "note": "Many embeddings are compact versions of bigger data." }, "feature-selection": { "label": "differs from …", "note": "Dimensionality reduction makes new features. Feature Selection picks old ones." } } }, "zh": { "fullName": "降维", "factExplain": "把高维数据映射成更少维度的紧凑表示。", "humanExplain": "拍证件照前，摄影师会喊收下巴抬肩膀：人还是那个人，但一下就从松散路人脸变清爽标准照。\n\n它常用于可视化、去噪和加速建模，也方便看清数据结构。", "humanExplainDisplay": "拍证件照前，\n摄影师会喊\n==收下巴抬肩膀==：\n人还是那个人，\n但一下就从松散路人脸\n变==清爽标准照==。\n\n它常用于可视化、去噪\n和加速建模，\n也方便看清数据结构。", "relationsNarrative": "Curse-of-dimensionality\n降维常用来缓解高维空间带来的稀疏与计算困难。\n\nPrincipal-component-analysis\nPCA 是最经典、最常见的降维方法之一。\n\nEmbedding\n很多嵌入方法，本质是在学更紧凑的数据表示。\n\nFeature-selection\n降维会组合出新特征，不只是删掉原特征。", "relations": { "curse-of-dimensionality": { "label": "缓解…难题", "note": "维度太高时，降维常是补救手段。" }, "principal-component-analysis": { "label": "经典方法是…", "note": "PCA 是最常见的线性降维方法。" }, "embedding": { "label": "压成…表示", "note": "很多嵌入本质上是在做紧凑表示。" }, "feature-selection": { "label": "不同于…取舍", "note": "它造新坐标，后者挑原特征。" } } } } }, { "id": "distillation", "name": "Distillation", "aliases": [ "知识蒸馏", "KD" ], "layer": "L2", "era": "2015", "publishedAt": "2026-05-23T09:35:00Z", "relations": [ { "to": "foundation-model" }, { "to": "fine-tuning" }, { "to": "local-llm" } ], "i18n": { "en": { "fullName": "Model Distillation", "factExplain": "A training method for moving big-model skills into a smaller model.", "humanExplain": "Distillation is like squeezing a huge cookbook onto a fridge magnet. You lose the tiny tips, but dinner starts faster.\n\nIt makes smaller AI models for phones and laptops. You trade a little skill for speed.", "humanExplainDisplay": "Distillation is like squeezing\na ==huge cookbook== onto\na ==fridge magnet==.\nYou lose the tiny tips,\nbut dinner starts faster.\n\nIt makes smaller AI models\nfor phones and laptops.\nYou trade a little skill\nfor speed.", "relationsNarrative": "Foundation-model\nDistillation squeezes a Foundation-model's skills into a smaller model.\n\nFine-tuning\nDistillation can keep special skills after Fine-tuning.\n\nLocal-LLM\nDistillation makes a Local-LLM need less local hardware.", "relations": { "foundation-model": { "label": "copies skills from …", "note": "Distillation squeezes Foundation-model skills into a smaller model." }, "fine-tuning": { "label": "can work after …", "note": "Distillation can keep the special skills learned by Fine-tuning." }, "local-llm": { "label": "makes … easier to run", "note": "Distillation lowers the hardware needs for a Local-LLM." } } }, "zh": { "fullName": "模型蒸馏", "factExplain": "把大模型能力压缩迁移到较小模型中的训练方法。", "humanExplain": "模型蒸馏像考前学霸给同桌划重点：不背整本书，只学最会考的解法。\n\n它常把大模型能力压进小模型，适合低成本部署和本地运行。", "humanExplainDisplay": "模型蒸馏像==考前学霸给同桌划重点==：\n==不背整本书==，\n只学最会考的解法。\n\n它常把大模型能力压进小模型，\n适合低成本部署和本地运行。", "relationsNarrative": "Foundation-model\nDistillation 将 Foundation-model 的能力压缩到小模型中。\n\nFine-tuning\nFine-tuning 后的专项能力，可通过 Distillation 保留下来。\n\nLocal-LLM\nDistillation 降低了 Local-LLM 对本地硬件的要求。", "relations": { "fine-tuning": { "label": "可配合…" }, "local-llm": { "label": "让…易部署" } } } } }, { "id": "distributed-computing", "name": "Distributed Computing", "layer": "L5", "sublayer": "architecture", "era": "1960s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "data-parallelism" }, { "to": "model-parallelism" }, { "to": "gpu" }, { "to": "ai-data-center" } ], "track": "history", "i18n": { "en": { "fullName": "Distributed Computing", "factExplain": "A way to split one big computing job across many machines that work together.", "humanExplain": "Distributed computing is a school cafeteria at lunch rush. One cook would melt, so the whole crew saves the pizza line.\n\nIt splits a huge job across many computers. You meet it in big AI training. It also powers giant data work and busy online services.", "humanExplainDisplay": "Distributed computing is a ==school cafeteria at lunch rush==.\nOne cook would melt,\nso the ==whole crew== saves the pizza line.\n\nIt splits a huge job\nacross many computers.\nYou meet it in big AI training.\nIt also powers giant data work\nand busy online services.", "relationsNarrative": "Data Parallelism\nData Parallelism is the most common way to split training work in distributed computing.\n\nModel parallelism\nModel parallel shares a model across devices when one machine cannot fit it.\n\nGPU\nDistributed computing often makes many GPUs work as one team.\n\nAI data center\nLarge distributed computing usually runs inside an AI data center.", "relations": { "data-parallelism": { "label": "splits data with …", "note": "Data Parallelism is a common way to divide distributed training." }, "model-parallelism": { "label": "splits models with …", "note": "Model parallel spreads a huge model when one machine cannot fit it." }, "gpu": { "label": "organizes … together", "note": "Distributed computing often links many GPUs into one team." }, "ai-data-center": { "label": "runs in …", "note": "Large distributed jobs usually run in data centers." } } }, "zh": { "fullName": "分布式计算（Distributed Computing）", "factExplain": "把计算任务拆到多台机器协同完成。", "humanExplain": "分布式计算像公司赶双十一大促：文案、客服、仓库、配送一起开工，单靠一人早就忙到宕机。\n\n它把大任务拆给多机协作，常用于大模型训练、海量数据处理和高并发服务。", "humanExplainDisplay": "分布式计算像公司赶双十一大促：\n文案、客服、仓库、\n配送==一起开工==，\n单靠一人早就忙到==宕机==。\n\n它把大任务拆给多机协作，\n常用于大模型训练、\n海量数据处理和高并发服务。", "relationsNarrative": "Data-parallelism\n数据并行是分布式计算里最常见的拆任务方式。\n\nModel parallelism\n当单机放不下模型时，会用它跨设备分摊。\n\nGPU\n分布式计算常把多张 GPU 组织起来一起干活。\n\nAI data center\n大规模分布式计算通常部署在数据中心内。", "relations": { "data-parallelism": { "label": "常用…拆活", "note": "数据并行是分布式训练常见做法。" }, "model-parallelism": { "label": "用…分模型", "note": "模型太大时可拆到多机协作。" }, "gpu": { "label": "组织…协作", "note": "常把多张 GPU 连成一套算力。" }, "ai-data-center": { "label": "跑在…里", "note": "大规模分布式计算通常落在数据中心。" } } } } }, { "id": "distributional-semantics", "name": "Dist. Semantics", "layer": "L2", "era": "1957", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "embedding" }, { "to": "word2vec" }, { "to": "representation-learning" }, { "to": "natural-language-processing" } ], "track": "history", "i18n": { "en": { "fullName": "Distributional Semantics", "factExplain": "A method that models word meaning from the words around it.", "humanExplain": "Words are like kids in a cafeteria. Sit at the same lunch table every day, and people guess your crowd.\n\nDistributional semantics uses nearby words to guess meaning. It helps with word vectors and search in NLP.", "humanExplainDisplay": "Words are like ==kids in a cafeteria==.\nSit at the ==same lunch table== every day,\nand people guess your crowd.\n\nDistributional semantics uses nearby words\nto guess meaning.\nIt helps with word vectors\nand search in NLP.", "relationsNarrative": "Embedding\nDistributional Semantics helped turn words into vector meanings.\n\nWord2Vec\nWord2Vec is one classic way to use Distributional Semantics.\n\nRepresentation Learning\nDistributional Semantics learns useful meaning patterns from text.\n\nNLP\nDistributional Semantics helps build text search and understanding.", "relations": { "embedding": { "label": "leads to …", "note": "It turns word meaning into numbers a computer can use." }, "word2vec": { "label": "becomes …", "note": "Word2Vec is a classic way to use this idea." }, "representation-learning": { "label": "belongs to …", "note": "It learns useful meaning patterns from text." }, "natural-language-processing": { "label": "supports …", "note": "It helps NLP compare, search, and understand text." } } }, "zh": { "fullName": "分布语义学", "factExplain": "根据词语在上下文中的共现分布来刻画词义。", "humanExplain": "词义像班里认同学：总一起值日、打球、挨老师点名的，十有八九就是一路人。\n\n它用于词表示和语义检索，帮系统靠上下文判断词义。", "humanExplainDisplay": "词义像班里认同学：\n总一起==值日、打球==、\n挨老师点名的，\n十有八九就是==一路人==。\n\n它用于词表示\n和语义检索，\n帮系统靠上下文判断词义。", "relationsNarrative": "Embedding\n它推动了词语从符号变成向量表示。\n\nWord2Vec\nWord2Vec 是分布语义学最经典的实现之一。\n\nRepresentation Learning\n它本质上是在学习可用的语义表示。\n\nNatural Language Processing\n它为文本相似度、检索和理解打基础。", "relations": { "embedding": { "label": "催生…表示", "note": "它把词义转成可计算的向量表示。" }, "word2vec": { "label": "落地成…方法", "note": "Word2Vec 是它的经典实现路线。" }, "representation-learning": { "label": "属于…思路", "note": "它关心如何学到有用语义表示。" }, "natural-language-processing": { "label": "服务于…任务", "note": "它长期是自然语言处理的基础方法。" } } } } }, { "id": "do-calculus", "name": "Do-Calculus", "layer": "L2", "era": "1995", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "structural-causal-model" }, { "to": "regression" }, { "to": "ai-bias" } ], "track": "history", "i18n": { "en": { "fullName": "Do-Calculus", "factExplain": "A rule system for working out what happens after a real intervention.", "humanExplain": "Do-Calculus is the rulebook for actually pressing the button. No staring at the elevator panel and blaming vibes.\n\nYou meet it in causal studies and decision checks. It asks what changes after we truly change an action.", "humanExplainDisplay": "Do-Calculus is the rulebook for\n==actually pressing the button==.\nNo staring at the elevator panel\nand ==blaming vibes==.\n\nYou meet it in causal studies\nand decision checks.\nIt asks what changes\nafter we truly change an action.", "relationsNarrative": "Structural Causal Model\nDo-Calculus often relies on a Structural Causal Model to define variables and interventions.\n\nRegression\nRegression is good at finding links, but it cannot directly answer what happens after an intervention.\n\nAI-bias\nDo-Calculus can help find where bias comes from and what truly causes it.", "relations": { "structural-causal-model": { "label": "builds on …", "note": "Do-Calculus often uses a Structural Causal Model to define interventions." }, "regression": { "label": "fixes limits of …", "note": "Regression can show links, but not always causes." }, "ai-bias": { "label": "helps trace sources of …", "note": "Do-Calculus can separate bias from mixing things up or changing actions." } } }, "zh": { "fullName": "do 演算 / 干预演算", "factExplain": "推导干预效应的因果规则系统。", "humanExplain": "do-calculus像门诊调药：这次真减一味药，再看症状怎么变，而不是拍脑袋猜病根。\n\n常用于因果分析和决策评估，判断真改动作后会怎样。", "humanExplainDisplay": "do-calculus像门诊调药：\n这次真减==一味药==，\n再看症状怎么变，\n而不是==拍脑袋猜病根==。\n\n常用于因果分析和决策评估，\n判断真改动作后会怎样。", "relationsNarrative": "Structural Causal Model\n它通常依赖结构因果模型来定义变量关系与干预。\n\nRegression\n回归擅长找相关性，但不能直接回答干预后会怎样。\n\nAI-bias\n它可用于分析偏差成因，区分相关偏差与因果影响。", "relations": { "structural-causal-model": { "label": "建立在…之上", "note": "它通常在结构因果模型里使用。" }, "regression": { "label": "纠正…局限", "note": "回归能看相关，未必能看因果。" }, "ai-bias": { "label": "帮助分析…来源", "note": "可拆分偏差来自混杂还是干预。" } } } } }, { "id": "document-parsing", "name": "Document parsing", "layer": "L5", "sublayer": "architecture", "era": "2010s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "rag" }, { "to": "structured-output" }, { "to": "multimodal" }, { "to": "llmops" } ], "track": "today", "i18n": { "en": { "fullName": "Document parsing", "factExplain": "The process of pulling document content into a structure software can use.", "humanExplain": "Doc parsing is the cafeteria tray return for messy files. A messy PDF goes in. Clean facts come out in neat bins.\n\nIt often comes before a knowledge base. It also helps contract and receipt bots.", "humanExplainDisplay": "Doc parsing is the ==cafeteria tray return==\nfor messy files.\nA messy PDF goes in.\nClean facts come out in ==neat bins==.\n\nIt often comes before a knowledge base.\nIt also helps contract and receipt bots.", "relationsNarrative": "RAG\nDoc parsing cleans the document first, so RAG can find the right content.\n\nStructured output\nDoc parsing turns pulled content into fields for the next program.\n\nMultimodal AI\nDoc parsing may use Multimodal skills to read scans and screenshots.\n\nLLMOps\nDoc parsing often sits in the data loading step for later AI apps.", "relations": { "rag": { "label": "feeds clean pages to …", "note": "Clean pages help RAG find the right content." }, "structured-output": { "label": "outputs as …", "note": "Parsed content often becomes named fields." }, "multimodal": { "label": "reads image pages with …", "note": "Scans need image understanding before text can be pulled out." }, "llmops": { "label": "plugs into …", "note": "It is often part of the data loading flow." } } }, "zh": { "fullName": "文档解析", "factExplain": "把文档内容提取并转成可处理结构的过程。", "humanExplain": "拍糊的合同、歪扭的扫描件，到它手里就像遇上强迫症管家：一张张抚平、按栏目归位，再乱也给你码成表格。\n\n常用于知识库入库、合同和单据处理，是自动化流程的起手式。", "humanExplainDisplay": "拍糊的合同、歪扭的扫描件，\n到它手里就像遇上==强迫症管家==：\n一张张抚平、按栏目归位，\n再乱也给你==码成表格==。\n\n常用于知识库入库、\n合同和单据处理，\n是自动化流程的起手式。", "relationsNarrative": "RAG\n文档先解析干净，RAG 才更容易检到对的内容。\n\nStructured output\n解析后的内容常要整理成字段，便于程序继续处理。\n\nMultimodal AI\n遇到扫描件、截图和图文混排时，常要靠多模态能力。\n\nLLMOps\n它常接在数据清洗和入库环节，服务后续应用。", "relations": { "rag": { "label": "给…喂资料", "note": "先把文档拆干净，检索才更靠谱。" }, "structured-output": { "label": "转成…格式", "note": "解析后常要输出字段化结果。" }, "multimodal": { "label": "处理图片页", "note": "遇到扫描件时常要看图识内容。" }, "llmops": { "label": "接入…流程", "note": "它常是数据入库链路的一环。" } } } } }, { "id": "domain-specific-slm", "name": "Domain-specific SLM", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "small-language-model" }, { "to": "fine-tuning" }, { "to": "distillation" }, { "to": "enterprise-ai-deployment" } ], "track": "today", "i18n": { "en": { "fullName": "Domain-specific Small Language Model", "factExplain": "A small language model tuned for one specific field.", "humanExplain": "It is the pancake cook of AI. No sushi. No fancy soufflé. Just great pancakes.\n\nYou meet it in finance, health care, and support. It handles expert tasks cheaply and steadily.", "humanExplainDisplay": "It is the ==pancake cook== of AI.\nNo sushi.\nNo fancy soufflé.\nJust ==great pancakes==.\n\nYou meet it in finance, health care, and support.\nIt handles expert tasks cheaply and steadily.", "relationsNarrative": "SLM\nA Domain-specific SLM is an SLM focused on one field.\n\nFine-tuning\nFine-tuning often turns a general model toward industry tasks.\n\nDistillation\nDistillation can squeeze big-model skill into a smaller model.\n\nEnterprise AI Deployment\nIt fits company use cases with low cost and more control.", "relations": { "small-language-model": { "label": "is a type of …", "note": "It is an SLM narrowed for one field." }, "fine-tuning": { "label": "adapts through …", "note": "Fine-tuning pushes a general model toward expert tasks." }, "distillation": { "label": "compresses skill with …", "note": "Distillation can pack big-model skill into a smaller expert." }, "enterprise-ai-deployment": { "label": "supports …", "note": "Companies often want AI that is cheap, useful, and controlled." } } }, "zh": { "fullName": "领域专用小语言模型", "factExplain": "针对特定领域优化的小语言模型。", "humanExplain": "领域专用小模型是社区老中医：不聊星辰大海，专把一种毛病看准。\n\n适合金融、医疗、客服，用低成本跑稳专业任务。", "humanExplainDisplay": "领域专用小模型是\n==社区老中医==：\n不聊星辰大海，\n专把==一种毛病看准==。\n\n适合金融、医疗、客服，\n用低成本跑稳专业任务。", "relationsNarrative": "SLM\n领域专用小模型是 SLM 的垂直化版本。\n\nFine-tuning\nFine-tuning 常把通用模型调向行业任务。\n\nDistillation\nDistillation 可把大模型能力压进更小模型。\n\nEnterprise AI Deployment\n它适合企业在低成本、可控场景落地。", "relations": { "small-language-model": { "label": "属于…", "note": "它是在小模型上做领域取舍。" }, "fine-tuning": { "label": "通过…适配领域", "note": "微调把通用能力拧向专业任务。" }, "distillation": { "label": "借…压缩能力", "note": "大模型知识可压进小专家。" }, "enterprise-ai-deployment": { "label": "支撑…", "note": "行业落地更看重够用和可控。" } } } } }, { "id": "doubao-2-1", "name": "Doubao 2.1", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "reasoning-model" }, { "to": "multimodal" }, { "to": "foundation-model" } ], "track": "today", "i18n": { "en": { "fullName": "Doubao 2.1", "factExplain": "A general AI model version in ByteDance’s Doubao family.", "humanExplain": "Doubao 2.1 is the brainy friend in the group chat. It reads the meme. Then it explains the scary math.\n\nYou meet it in chat, writing, search, and image tasks. It is the base model for the Doubao world.", "humanExplainDisplay": "Doubao 2.1 is the ==brainy friend==\nin the group chat.\nIt ==reads the meme==.\nThen it explains the scary math.\n\nYou meet it in chat, writing, search,\nand image tasks.\nIt is the base model\nfor the Doubao world.", "relationsNarrative": "LLM\nDoubao 2.1 is a general LLM for chat, writing, and code.\n\nReasoning-model\nDoubao 2.1 uses stronger reasoning to handle harder questions.\n\nMultimodal AI\nDoubao 2.1 can support text and image understanding.\n\nFoundation-model\nDoubao 2.1 can serve as the base model for Doubao apps and APIs.", "relations": { "llm": { "label": "is a kind of …", "note": "It is an LLM built for general tasks." }, "reasoning-model": { "label": "improves … skills", "note": "This version focuses on harder reasoning tasks." }, "multimodal": { "label": "supports … understanding", "note": "It can grow into text and image understanding." }, "foundation-model": { "label": "serves as a …", "note": "It can power higher-level Doubao apps and APIs." } } }, "zh": { "fullName": "豆包 2.1", "factExplain": "字节跳动豆包系列的通用大模型版本。", "humanExplain": "豆包 2.1 像群里那个学霸朋友：接得住梗，也能把难题讲明白。\n\n用于聊天、写作、搜索和图文理解，是豆包生态底座。", "humanExplainDisplay": "豆包 2.1 像群里那个学霸朋友：\n接得住梗，\n也能把==难题讲明白==。\n\n用于聊天、写作、搜索\n和图文理解，\n是豆包生态底座。", "relationsNarrative": "LLM\n豆包 2.1 是面向对话、写作、代码的通用 LLM。\n\nReasoning Model\n它通过推理能力，处理更复杂的问题。\n\nMultimodal AI\n它可接入图文理解，服务更丰富场景。\n\nFoundation-model\n它可作为豆包应用和 API 的基础模型。", "relations": { "llm": { "label": "属于…", "note": "它是面向通用任务的大语言模型。" }, "reasoning-model": { "label": "提升…能力", "note": "新版重点提升复杂推理能力。" }, "multimodal": { "label": "支持…理解", "note": "可扩展到图文等多模态理解。" }, "foundation-model": { "label": "充当…底座", "note": "它可作为上层应用的模型底座。" } } } } }, { "id": "double-descent", "name": "Double Descent", "layer": "L2", "era": "2019", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "overparameterization" }, { "to": "bias-variance-tradeoff" }, { "to": "scaling-law" }, { "to": "statistical-learning-theory" } ], "track": "history", "i18n": { "en": { "fullName": "Double Descent", "factExplain": "A model's test mistakes can drop, rise, then drop again as it gets bigger.", "humanExplain": "Double Descent is like upgrading a toy car. One button bumps the couch. A few buttons crash harder. Full controls finally drive straight.\n\nIt shows test mistakes can fall, rise, then fall again as a model gets bigger. In large model tests, the old U-shaped curve can fool you.", "humanExplainDisplay": "Double Descent is like ==upgrading a toy car==.\nOne button bumps the couch.\nA few buttons crash harder.\n==Full controls== finally drive straight.\n\nIt shows test mistakes can fall,\nrise,\nthen fall again\nas a model gets bigger.\nIn large model tests,\nthe old U-shaped curve can fool you.", "relationsNarrative": "Overparameterization\nDouble Descent often appears after a model has more parameters than examples.\n\nBias-Variance Tradeoff\nDouble Descent challenges the old U-shaped view of test mistakes.\n\nScaling-law\nDouble Descent shows bigger models can sometimes still get better.\n\nSLT\nDouble Descent pushes SLT to explain generalization again.", "relations": { "overparameterization": { "label": "often appears with …", "note": "Overparameterization can let test mistakes drop again." }, "bias-variance-tradeoff": { "label": "challenges …", "note": "Double Descent makes the classic U-shaped curve too simple." }, "scaling-law": { "label": "echoes …", "note": "Making a model bigger can still make it better." }, "statistical-learning-theory": { "label": "pushes …", "note": "It pushes theory to explain why big models can still generalize." } } }, "zh": { "fullName": "双重下降", "factExplain": "测试误差随模型复杂度先降、升、再降的现象。", "humanExplain": "双重下降像考研刷题：题太少不会，刚会背最飘；题海刷穿，反而摸到套路。\n\n提示大模型未必更差，评估泛化别只套 U 形。", "humanExplainDisplay": "双重下降像考研刷题：\n题太少不会，\n刚会==背最飘==；\n题海刷穿，反而==摸到套路==。\n\n提示大模型未必更差，\n评估泛化，\n别只套 U 形。", "relationsNarrative": "Overparameterization\n双重下降常出现在参数多于样本之后。\n\nBias-Variance Tradeoff\n它挑战了传统 U 形泛化误差直觉。\n\nScaling-law\n它说明继续变大有时仍能变好。\n\nStatistical Learning Theory\n它推动理论重新解释泛化。", "relations": { "overparameterization": { "label": "常见于…", "note": "过参数化让误差可能再次下降。" }, "bias-variance-tradeoff": { "label": "挑战…", "note": "它让经典 U 形曲线不再够用。" }, "scaling-law": { "label": "呼应…", "note": "继续放大模型有时仍会变好。" }, "statistical-learning-theory": { "label": "推动…", "note": "它促使泛化理论重新解释。" } } } } }, { "id": "dpll-algorithm", "name": "DPLL", "layer": "L2", "era": "1962", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "constraint-satisfaction-problem" }, { "to": "resolution-principle" }, { "to": "unification" } ], "track": "history", "i18n": { "en": { "fullName": "Davis–Putnam–Logemann–Loveland algorithm", "factExplain": "A classic backtracking algorithm for solving Boolean satisfiability puzzles.", "humanExplain": "DPLL is like doing a giant Sudoku with a bossy pencil. It tries a yes-or-no guess, then erases fast when the puzzle says “nope.”\n\nIt is used in SAT solvers and automated reasoning. It backtracks from conflicts to find true-or-false choices that fit.", "humanExplainDisplay": "DPLL is like doing a ==giant Sudoku==\nwith a ==bossy pencil==.\nIt tries a yes-or-no guess,\nthen erases fast when the puzzle says “nope.”\n\nIt is used in SAT solvers\nand automated reasoning.\nIt backtracks from conflicts\nto find true-or-false choices that fit.", "relationsNarrative": "CSP\nDPLL is close to CSP. Both look for choices that obey all rules.\n\nResolution\nDPLL uses backtracking search. Resolution uses logical proof.\n\nUnification\nDPLL and Unification are both basic tools in classic symbolic AI.", "relations": { "constraint-satisfaction-problem": { "label": "solves a cousin of …", "note": "Both search for choices that satisfy all rules." }, "resolution-principle": { "label": "is often compared with …", "note": "DPLL uses search. Resolution uses logic steps." }, "unification": { "label": "shares a toolbox with …", "note": "Both are basic tools in classic symbolic AI." } } }, "zh": { "fullName": "戴维斯-普特南-洛夫兰-洛格曼算法", "factExplain": "一种用于求解布尔可满足性问题的经典回溯算法。", "humanExplain": "它像宿舍查违禁品：先翻一个柜子，若整层楼都对不上，就顺着线索倒回去重查，直到每间都自洽。\n\n常用于 SAT 求解和自动推理，在冲突中回溯找可行赋值。", "humanExplainDisplay": "它像宿舍查违禁品：\n先翻一个柜子，\n若整层楼都==对不上==，\n就倒回去==重查==。\n\n常用于 SAT 求解\n和自动推理，\n在冲突中回溯找可行赋值。", "relationsNarrative": "Constraint Satisfaction Problem\n它和约束满足问题很近，都是在一堆限制里找可行解。\n\nResolution Principle\n它常和归结法对照：一个靠回溯搜索，一个靠逻辑推出矛盾。\n\nUnification\n它和合一都属于经典符号推理里的基础工具。", "relations": { "constraint-satisfaction-problem": { "label": "求解…的近亲", "note": "两者都在找满足约束的解。" }, "resolution-principle": { "label": "常与…对照", "note": "一个偏搜索回溯，一个偏逻辑推导。" }, "unification": { "label": "同属符号推理工具箱", "note": "都常见于经典 AI 与自动证明。" } } } } }, { "id": "dropout", "name": "Dropout", "layer": "L2", "era": "2012", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "neural-network" }, { "to": "weight-decay" }, { "to": "alexnet" } ], "track": "history", "i18n": { "en": { "fullName": "Dropout", "factExplain": "A training trick that randomly turns off some neurons to reduce overfitting.", "humanExplain": "Dropout is like a coach benching random players at practice. Now everyone must learn the play.\n\nIn neural-network training, it turns off some neurons for a moment. This helps the model avoid overfitting and plain old memorizing.", "humanExplainDisplay": "Dropout is like a coach\n==benching random players== at practice.\nNow ==everyone must learn the play==.\n\nIn neural-network training,\nit turns off some neurons for a moment.\nThis helps the model avoid overfitting\nand plain old memorizing.", "relationsNarrative": "Regularization\nDropout is a Regularization method. It randomly turns neurons off to fight overfitting.\n\nNeural-network\nDropout turns off some neurons during Neural-network training.\n\nWeight Decay\nBoth fight overfitting: Dropout turns neurons off, Weight Decay keeps weights small.\n\nAlexNet\nAlexNet used Dropout to reduce overfitting and made it famous.", "relations": { "regularization": { "label": "is a kind of …", "note": "It fights overfitting by randomly turning neurons off." }, "neural-network": { "label": "turns off nodes in …", "note": "During training, it hides some neurons for a short time." }, "weight-decay": { "label": "fights overfitting with …", "note": "Dropout turns neurons off. Weight Decay keeps weights small." }, "alexnet": { "label": "was popularized by …", "note": "AlexNet used Dropout to reduce overfitting and made it famous." } } }, "zh": { "fullName": "随机失活", "factExplain": "训练时随机屏蔽部分神经元的正则化方法。", "humanExplain": "Dropout 像课堂随机禁学霸：别全靠尖子答题，逼全班都得会。\n\n用于神经网络训练，减少过拟合和死记硬背。", "humanExplainDisplay": "Dropout 像课堂\n==随机禁学霸==：\n别全靠尖子答题，\n逼==全班都得会==。\n\n用于神经网络训练，\n减少过拟合，\n和死记硬背。", "relationsNarrative": "Regularization\nDropout 是正则化方法，用随机屏蔽防过拟合。\n\nNeural-network\n它训练神经网络时，临时关掉部分神经元。\n\nWeight Decay\n两者都防过拟合，一个关节点，一个压权重。\n\nAlexNet\nAlexNet 使用它缓解过拟合，并带火深度学习。", "relations": { "regularization": { "label": "属于…", "note": "它用随机关节点防止过拟合。" }, "neural-network": { "label": "在…中关节点", "note": "训练时临时屏蔽部分神经元。" }, "weight-decay": { "label": "和…同防过拟合", "note": "一个关神经元，一个压权重。" }, "alexnet": { "label": "被…带火", "note": "AlexNet 用它缓解过拟合。" } } } } }, { "id": "dyna", "name": "Dyna", "layer": "L2", "era": "1990", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "model-based-reinforcement-learning" }, { "to": "temporal-difference-learning" }, { "to": "q-learning" } ], "track": "history", "seo": { "zh": { "title": "Dyna 强化学习框架 是什么?武侠学徒的脑内擂台,一文看懂 — AI Rookies", "description": "把真实经验和模型模拟一起用于强化学习。人话版一分钟讲明白。" }, "en": { "title": "What Is Dyna? Explained in Plain English", "description": "A reinforcement learning framework using both real experience and model-made practice. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Dyna Reinforcement Learning Framework", "factExplain": "A reinforcement learning framework using both real experience and model-made practice.", "humanExplain": "Dyna is like driver's ed with a simulator. The student drives one real lap, then does ten pretend laps without hitting real cones.\n\nIt helps an agent learn with fewer real-world mistakes. You see it in MBRL and robot control.", "humanExplainDisplay": "Dyna is like ==driver's ed with a simulator==.\nThe student drives one real lap,\nthen does ==ten pretend laps==\nwithout hitting real cones.\n\nIt helps an agent learn\nwith fewer real-world mistakes.\nYou see it in MBRL\nand robot control.", "relationsNarrative": "MBRL\nDyna is an early MBRL framework for planning with a learned model.\n\nTD Learning\nDyna uses TD Learning on both real and simulated experience.\n\nQ-Learning\nDyna-Q adds planning steps to Q-Learning.", "relations": { "model-based-reinforcement-learning": { "label": "is part of …", "note": "Dyna is an early example of MBRL." }, "temporal-difference-learning": { "label": "updates values with …", "note": "Dyna can run TD updates on real and simulated experience." }, "q-learning": { "label": "extends …", "note": "Dyna-Q adds planning to Q-Learning." } } }, "zh": { "fullName": "Dyna 强化学习框架", "factExplain": "把真实经验和模型模拟一起用于强化学习。", "humanExplain": "Dyna像武侠学徒练功：白天真过招挨打，夜里在脑内擂台再打一百遍。\n\n让智能体少试错多练习，常用于模型式强化学习和机器人控制。", "humanExplainDisplay": "Dyna像武侠学徒练功：\n白天==真过招挨打==，\n夜里在==脑内擂台==\n再打一百遍。\n\n让智能体少试错多练习，\n常用于模型式强化学习，\n和机器人控制。", "relationsNarrative": "MBRL\nDyna 是早期把学习模型用于规划的代表框架。\n\nTD Learning\nDyna 常用 TD 更新真实和模拟经验。\n\nQ-Learning\nDyna-Q 把规划步骤接到 Q-Learning 上。", "relations": { "model-based-reinforcement-learning": { "label": "属于…", "note": "Dyna 是模型式强化学习的早期代表。" }, "temporal-difference-learning": { "label": "用…更新价值", "note": "真实和模拟经验都可做 TD 更新。" }, "q-learning": { "label": "扩展…", "note": "Dyna-Q 把规划接到 Q-Learning 上。" } } } } }, { "id": "dynamic-programming", "name": "DP", "layer": "L2", "era": "1957", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "bellman-equation" }, { "to": "value-iteration" }, { "to": "policy-iteration" }, { "to": "q-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Dynamic Programming", "factExplain": "A way to split a big problem into repeat parts and reuse saved answers.", "humanExplain": "Dynamic Programming is pancake math. Make one bowl of batter, not a new egg mess for every pancake.\n\nIt saves repeat answers, then uses them again. You meet it in best-choice or best-path tasks, and in many reinforcement learning methods.", "humanExplainDisplay": "Dynamic Programming is ==pancake math==.\nMake ==one bowl of batter==,\nnot a new egg mess for every pancake.\n\nIt saves repeat answers,\nthen uses them again.\nYou meet it in best-choice or best-path tasks,\nand in many reinforcement learning methods.", "relationsNarrative": "Bellman Equation\nDynamic Programming often uses the Bellman Eq to update answers step by step.\n\nValue Iteration\nValue Iteration uses Dynamic Programming to update values again and again.\n\nPolicy Iteration\nPolicy Iter. uses Dynamic Programming to rate a plan and improve it.\n\nQ-Learning\nQ-Learning borrows the idea, but it can learn without a full world map.", "relations": { "bellman-equation": { "label": "is often written as …", "note": "Dynamic Programming often uses the Bellman Eq for its repeat update." }, "value-iteration": { "label": "powers …", "note": "Value Iteration is a classic Dynamic Programming method." }, "policy-iteration": { "label": "supports … updates", "note": "Policy Iter. uses Dynamic Programming to rate a plan and improve it." }, "q-learning": { "label": "inspires …", "note": "Q-Learning borrows the idea, but it does not need the full world map first." } } }, "zh": { "fullName": "Dynamic Programming｜动态规划", "factExplain": "把大问题拆成重叠子问题并复用结果的方法。", "humanExplain": "像做糖葫芦先熬一锅糖浆：这串裹好了，后面同样的果子别再从白糖重新起火。\n\n常用于求最优解和路径，也是很多强化学习方法的基础。", "humanExplainDisplay": "像做糖葫芦先熬一锅糖浆：\n这串裹好了，后面同样的果子\n别再从==白糖==重新 ==起火==。\n\n常用于求最优解和路径，\n也是很多强化学习方法的基础。", "relationsNarrative": "Bellman-equation\n动态规划常把最优问题写成贝尔曼方程来递推。\n\nValue-iteration\n价值迭代是用动态规划反复更新价值的经典做法。\n\nPolicy-iteration\n策略迭代先评估再改进，核心也建立在动态规划上。\n\nQ-Learning\nQ 学习继承了它的思想，但不必先知道环境全貌。", "relations": { "bellman-equation": { "label": "常写成…", "note": "动态规划常靠贝尔曼方程递推。" }, "value-iteration": { "label": "用于…求解", "note": "价值迭代是动态规划的经典算法。" }, "policy-iteration": { "label": "支撑…更新", "note": "策略迭代靠它反复评估与改进。" }, "q-learning": { "label": "启发…设计", "note": "Q 学习像不完整环境下的近似版。" } } } } }, { "id": "early-stopping", "name": "Early Stopping", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "cross-validation" }, { "to": "gradient-descent" } ], "track": "history", "i18n": { "en": { "fullName": "Early Stopping", "factExplain": "Early Stopping stops training when validation scores start getting worse.", "humanExplain": "Early Stopping is pulling cookies out at golden brown. Wait for \"extra sure,\" and you make tiny hockey pucks.\n\nIn neural network training, it watches validation scores and stops when they slip. This saves compute and helps stop overfitting.", "humanExplainDisplay": "Early Stopping is pulling cookies out at ==golden brown==.\nWait for \"extra sure,\"\nand you make ==tiny hockey pucks==.\n\nIn neural network training,\nit watches validation scores\nand stops when they slip.\nThis saves compute\nand helps stop overfitting.", "relationsNarrative": "Regularization\nEarly Stopping works as Regularization by stopping early and reducing overfitting.\n\nCross-Validation\nCross-Validation gives validation scores for deciding when to stop training.\n\nGradient Descent\nEarly Stopping usually happens during Gradient Descent training.", "relations": { "regularization": { "label": "works as …", "note": "Early Stopping limits memorizing by stopping training early." }, "cross-validation": { "label": "relies on …", "note": "Validation scores often decide when training should stop." }, "gradient-descent": { "label": "stops …", "note": "It hits the brakes before more steps make results worse." } } }, "zh": { "fullName": "早停", "factExplain": "验证集性能变差时提前停止训练。", "humanExplain": "早停像健身教练喊停：刚练到位就收，别等动作变形还硬撑。\n\n用于神经网络训练，省算力，也减轻过拟合。", "humanExplainDisplay": "早停像健身教练喊停：\n==刚练到位就收==，\n别等动作变形\n还硬撑。\n\n用于神经网络训练，\n省算力，\n也减轻过拟合。", "relationsNarrative": "Regularization\n早停是一种通过提前停止来防止过拟合的正则化方法。\n\nCross-Validation\n验证集表现常被用来判断训练是否该停。\n\nGradient Descent\n早停通常发生在梯度下降等迭代训练过程中。", "relations": { "regularization": { "label": "作为…手段", "note": "早停用提前收手限制模型过度记忆。" }, "cross-validation": { "label": "依赖…判断", "note": "验证表现常用来决定何时停止训练。" }, "gradient-descent": { "label": "中止…过程", "note": "它在迭代优化变坏前踩刹车。" } } } } }, { "id": "edge-ai", "name": "Edge AI", "layer": "L5", "sublayer": "product", "era": "2010s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "ai-device-ai" }, { "to": "quantization" }, { "to": "model-compression" }, { "to": "local-llm" } ], "track": "today", "i18n": { "en": { "fullName": "Edge AI", "factExplain": "A way to run AI on local devices, not distant servers.", "humanExplain": "Edge AI is like putting a tiny bouncer inside your doorbell. It makes the call right there, not downtown.\n\nYou meet it in phones and cameras. Cars use it too. It reacts fast and sends less private data away.", "humanExplainDisplay": "Edge AI is like putting a ==tiny bouncer==\ninside your doorbell.\nIt ==makes the call right there==,\nnot downtown.\n\nYou meet it in phones and cameras.\nCars use it too.\nIt reacts fast.\nIt sends less private data away.", "relationsNarrative": "AI device\nAn AI device is often the hardware home for Edge AI.\n\nQuantization\nQuantization makes the model smaller for edge devices.\n\nModel Compression\nModel Compression lowers memory and compute pressure.\n\nLocal-LLM\nLocal-LLM is Edge AI for language tasks.", "relations": { "ai-device-ai": { "label": "runs on …", "note": "AI devices are common homes for Edge AI." }, "quantization": { "label": "shrinks with …", "note": "Quantization makes models small enough for tiny devices." }, "model-compression": { "label": "lightens with …", "note": "Model Compression cuts memory and compute needs at the edge." }, "local-llm": { "label": "includes … use cases", "note": "Local-LLM is Edge AI for language tasks." } } }, "zh": { "fullName": "边缘人工智能", "factExplain": "在本地边缘设备上运行 AI 推理的部署方式。", "humanExplain": "边缘 AI 像小区门口的社区门诊：头疼脑热当场看，别动不动挤三甲。\n\n常用于手机、摄像头、车机，响应快，也少上传隐私。", "humanExplainDisplay": "边缘 AI 像小区门口的\n==社区门诊==：\n头疼脑热当场看，\n别动不动挤==三甲==。\n\n常用于手机、摄像头、车机，\n响应快，\n也少上传隐私。", "relationsNarrative": "AI Device\nAI 设备常是边缘 AI 落地的硬件载体。\n\nQuantization\n量化把模型变小，方便在边缘端运行。\n\nModel Compression\n模型压缩减少算力和内存压力。\n\nLocal-LLM\n本地大模型是边缘 AI 在语言场景的分支。", "relations": { "ai-device-ai": { "label": "部署在…上", "note": "AI 设备是边缘 AI 常见载体。" }, "quantization": { "label": "靠…瘦身", "note": "量化让模型更适合小设备运行。" }, "model-compression": { "label": "借…减负", "note": "模型压缩降低边缘端算力压力。" }, "local-llm": { "label": "包含…场景", "note": "本地大模型是边缘部署的一类应用。" } } } } }, { "id": "electronic-skin", "name": "Electronic skin", "layer": "L5", "sublayer": "architecture", "era": "2000s", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "physical-ai-ai" }, { "to": "embodied-ai" }, { "to": "robotics" }, { "to": "dexterous-robotic-hand" } ], "track": "today", "i18n": { "en": { "fullName": "Electronic skin", "factExplain": "A flexible sensor system for machine touch.", "humanExplain": "Electronic skin is a robot’s smart sticker suit. Poke it or press it, and it tattles right away.\n\nIt senses touch and force. You meet it in prosthetic limbs and care robots. Wearables use it too.", "humanExplainDisplay": "Electronic skin is a robot’s ==smart sticker suit==.\nPoke it or press it,\nand it ==tattles right away==.\n\nIt senses touch and force.\nYou meet it in prosthetic limbs and care robots.\nWearables use it too.", "relationsNarrative": "Physical AI\nElectronic skin turns real touch into machine-ready signals.\n\nEmbodied AI\nEmbodied AI needs touch to understand the world more safely.\n\nRobotics\nElectronic skin gives robots a skin-like sense of touch.\n\nDexterous robotic hand\nA dexterous robotic hand uses it to judge grip force and slipping.", "relations": { "physical-ai-ai": { "label": "supports … sensing", "note": "It turns real touch into machine-ready signals." }, "embodied-ai": { "label": "gives touch to …", "note": "Body touch helps Embodied AI understand its surroundings." }, "robotics": { "label": "adds touch to …", "note": "Electronic skin lets robots feel contact." }, "dexterous-robotic-hand": { "label": "improves … grip", "note": "Fine touch helps a robotic hand control force." } } }, "zh": { "fullName": "电子皮肤", "factExplain": "一种模拟皮肤触觉的柔性传感系统。", "humanExplain": "电子皮肤像给机器人穿触觉秋衣：碰哪儿、压多重，立刻打小报告。\n\n用于假肢、护理机器人、穿戴设备，感知接触与力度。", "humanExplainDisplay": "电子皮肤像给机器人\n穿==触觉秋衣==：\n碰哪儿、压多重，\n立刻==打小报告==。\n\n用于假肢、护理机器人、穿戴设备，\n感知接触与力度。", "relationsNarrative": "Physical AI\n电子皮肤把真实接触转成机器可读信号。\n\nEmbodied AI\n具身智能需要触觉，才能更稳地理解环境。\n\nRobotics\n它给机器人补上类似皮肤的触觉感知。\n\nDexterous Robotic Hand\n灵巧手靠它判断抓握力度和滑动。", "relations": { "physical-ai-ai": { "label": "支撑…感知", "note": "把真实接触转成机器可读信号。" }, "embodied-ai": { "label": "为…提供触觉", "note": "身体触觉是具身智能的重要输入。" }, "robotics": { "label": "给…增加触觉", "note": "电子皮肤让机器人感知接触。" }, "dexterous-robotic-hand": { "label": "增强…抓握", "note": "细腻触觉帮助机械手控力度。" } } } } }, { "id": "eligibility-traces", "name": "Eligibility Traces", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "temporal-difference-learning" }, { "to": "reinforcement-learning" }, { "to": "actor-critic" } ], "track": "history", "i18n": { "en": { "fullName": "Eligibility Traces", "factExplain": "A reinforcement learning memory for giving recent moves credit after delayed rewards.", "humanExplain": "Eligibility traces are fading sticky notes on an AI's recent moves. A late reward reads the newest notes first. Old notes get tiny cookie crumbs.\n\nYou see them in TD Learning and Actor-Critic. They help late rewards travel back faster.", "humanExplainDisplay": "Eligibility traces are ==fading sticky notes==\non an AI's recent moves.\nA late reward reads the ==newest notes== first.\nOld notes get tiny cookie crumbs.\n\nYou see them in TD Learning and Actor-Critic.\nThey help late rewards travel back faster.", "relationsNarrative": "TD Learning\nEligibility traces let TD Learning give reward credit to recent steps.\n\nRL\nEligibility traces help RL learn from late rewards.\n\nActor-Critic\nActor-Critic can use eligibility traces to update its policy and value.", "relations": { "temporal-difference-learning": { "label": "extends credit in …", "note": "Eligibility traces let TD Learning look back at recent steps." }, "reinforcement-learning": { "label": "helps … with delayed rewards", "note": "They help an RL agent credit earlier moves." }, "actor-critic": { "label": "passes credit through …", "note": "They help Actor-Critic update the actor and the critic." } } }, "zh": { "fullName": "资格迹", "factExplain": "强化学习中记录近期状态动作信用的机制。", "humanExplain": "资格迹像接力赛记功：冲线的人亮眼，前几棒也按近远分功劳。\n\n用于 TD 和 Actor-Critic，让延迟奖励更快回传。", "humanExplainDisplay": "资格迹像接力赛记功：\n==冲线的人亮眼==，\n前几棒也按近远\n分功劳。\n\n用于 TD 和 Actor-Critic，\n让延迟奖励更快回传。", "relationsNarrative": "TD Learning\n资格迹让 TD 学习能把奖励分给近期步骤。\n\nRL\n资格迹帮助强化学习处理延迟奖励问题。\n\nActor-Critic\nActor-Critic 可用资格迹同时更新策略与价值。", "relations": { "temporal-difference-learning": { "label": "扩展…的信用分配", "note": "它让 TD 学习能回溯近期步骤。" }, "reinforcement-learning": { "label": "服务…的奖励学习", "note": "它帮助智能体处理延迟奖励。" }, "actor-critic": { "label": "为…传递信用", "note": "Actor-Critic 可用它更新策略与价值。" } } } } }, { "id": "eliza", "name": "ELIZA", "layer": "L5", "sublayer": "product", "era": "1966", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "chatbot" }, { "to": "natural-language-processing" }, { "to": "symbolic-ai" }, { "to": "turing-test" } ], "track": "history", "i18n": { "en": { "fullName": "ELIZA", "factExplain": "An early chat program using pattern matching and set replies.", "humanExplain": "ELIZA is like a school nurse with a clipboard. You say, “I feel stressed,” and it says, “Why do you feel stressed?”\n\nIt helps explain early chat systems. It shows a good reply is not real understanding.", "humanExplainDisplay": "ELIZA is like a ==school nurse with a clipboard==.\nYou say, “I feel stressed,”\nand it says, ==“Why do you feel stressed?”==\n\nIt helps explain early chat systems.\nIt shows a good reply\nis not real understanding.", "relationsNarrative": "Chatbot\nELIZA was an early chatbot and showed how easily chat can feel human.\n\nNLP\nELIZA used simple rules to handle human language input.\n\nSymbolic AI\nELIZA used hand-written rules and pattern matching to fake a chat.\n\nTuring-test\nELIZA shows a key lesson: replying well is not understanding.", "relations": { "chatbot": { "label": "pioneered early …", "note": "ELIZA was an early face of chatbots." }, "natural-language-processing": { "label": "showed early …", "note": "It used simple rules to handle human language." }, "symbolic-ai": { "label": "used … ideas", "note": "It made chat feel real with hand-written rules." }, "turing-test": { "label": "raises the … question", "note": "A smooth reply does not mean real understanding." } } }, "zh": { "fullName": "伊莱莎，早期聊天机器人", "factExplain": "一种基于模式匹配的早期聊天程序。", "humanExplain": "ELIZA 像相亲桌上的捧哏：抓你一句话反问，气氛像懂你，其实照稿走。\n\n用于理解早期对话系统；提醒会接话不等于会理解。", "humanExplainDisplay": "ELIZA 像\n==相亲桌上的捧哏==：\n抓你一句话反问，\n气氛像懂你，其实==照稿走==。\n\n用于理解早期对话系统；\n提醒会接话，\n不等于会理解。", "relationsNarrative": "Chatbot\n它是聊天机器人的早期代表，证明人会被对话感打动。\n\nNLP\n它用简单规则处理自然语言输入，是早期应用案例。\n\nSymbolic AI\n它靠人工规则和模式匹配，制造“像在聊天”的感觉。\n\nTuring Test\n它让人看到：会接话，不等于真的理解。", "relations": { "chatbot": { "label": "开创…雏形", "note": "它是聊天机器人的早期代表。" }, "natural-language-processing": { "label": "展示…早期应用", "note": "它用简单规则处理自然语言。" }, "symbolic-ai": { "label": "采用…思路", "note": "它靠人工规则制造对话感。" }, "turing-test": { "label": "呼应…问题", "note": "会接话，不等于真理解。" } } } } }, { "id": "embedding", "name": "Embedding", "layer": "L2", "era": "2013", "publishedAt": "2026-05-23T08:20:00Z", "relations": [ { "to": "vector-db" }, { "to": "rag" }, { "to": "neural-network" } ], "i18n": { "en": { "fullName": "Vector Embedding", "factExplain": "A way to turn text or images into numbers a computer can compare.", "humanExplain": "Embedding is like a seating chart for the school cafeteria. Things with the same vibe sit near each other.\n\nYou meet it in search and recommendations. RAG uses it too. It helps AI find by meaning, not just matching words.", "humanExplainDisplay": "Embedding is like a ==seating chart==\nfor the school cafeteria.\nThings with the ==same vibe==\nsit near each other.\n\nYou meet it in search and recommendations.\nRAG uses it too.\nIt helps AI find by meaning,\nnot just matching words.", "relationsNarrative": "Vector-db\nA Vector-db uses embeddings to search by meaning, not just words.\n\nRAG\nRAG uses embeddings to find related content in a knowledge base.\n\nNeural-network\nA Neural-network learns patterns and makes embeddings.", "relations": { "vector-db": { "label": "is stored in …", "note": "A Vector-db stores embeddings for meaning-based search." }, "rag": { "label": "powers search for …", "note": "RAG uses embeddings to find useful source text." }, "neural-network": { "label": "is made by …", "note": "A Neural-network learns patterns and makes embeddings." } } }, "zh": { "fullName": "向量嵌入", "factExplain": "把文本、图片等内容转换成可比较的数字向量。", "humanExplain": "它像给句子安排小区门牌：意思越像，住得越近，串门也快。\n\n它支撑语义搜索、推荐和知识库问答，让系统先找对材料再回答。", "humanExplainDisplay": "它像给句子安排==小区门牌==：\n意思越像，住得越近，\n==串门也快==。\n\n它支撑语义搜索、\n推荐和知识库问答，\n让系统先找对材料再回答。", "relationsNarrative": "Vector-db\nVector-db 依赖 Embedding 才能按语义相似度检索。\n\nRAG\nRAG 使用 Embedding 在资料库中找到相关内容。\n\nNeural-network\nNeural-network 训练出的表示能力，是 Embedding 的基础。", "relations": { "vector-db": { "label": "被…存储" }, "rag": { "label": "支撑…检索" }, "neural-network": { "label": "由…生成" } } } } }, { "id": "embodied-ai", "name": "Embodied AI", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "computer-use" }, { "to": "world-model" }, { "to": "agent" } ], "track": "today", "i18n": { "en": { "fullName": "Embodied AI", "factExplain": "AI that can sense and act in the physical world.", "humanExplain": "Normal AI gives advice from the couch. Embodied AI gets up and bumps the coffee table.\n\nIt senses the real world and acts in it. You see it in robots and self-driving cars. It also shows up in smart devices.", "humanExplainDisplay": "Normal AI gives ==advice from the couch==.\nEmbodied AI ==gets up==\nand bumps the coffee table.\n\nIt senses the real world\nand acts in it.\nYou see it in robots\nand self-driving cars.\nIt also shows up\nin smart devices.", "relationsNarrative": "Multimodal AI\nEmbodied AI uses Multimodal input to sense the world.\n\nComputer use\nComputer use works on screens, but Embodied AI steps into the real world.\n\nWorld model\nA World model helps it understand the scene and predict what happens next.\n\nAgent\nEmbodied AI gives an Agent a body for real-world action.", "relations": { "multimodal": { "label": "connects senses to …", "note": "It uses Multimodal input to sense the world." }, "computer-use": { "label": "goes beyond …", "note": "Computer use handles screens, but Embodied AI handles the real world." }, "world-model": { "label": "uses … to predict the room", "note": "A World model helps it predict what will happen next." }, "agent": { "label": "gives … a body", "note": "Embodied AI gives a planning Agent hands." } } }, "zh": { "fullName": "具身智能", "factExplain": "能通过身体感知并作用于物理世界的 AI。", "humanExplain": "这不是纸上谈兵的军师，而是能自己看路、出招、搬东西的下场选手。\n\n常见于机器人、自动驾驶和智能设备，让 AI 直接在现实世界行动。", "humanExplainDisplay": "这不是\n==纸上谈兵的军师==，\n而是能自己看路、出招、\n搬东西的\n==下场选手==。\n\n常见于机器人、\n自动驾驶和智能设备，\n让 AI 直接在现实世界行动。", "relationsNarrative": "Multimodal AI\n具身智能通常要融合视觉、语音和触觉等多模态信息。\n\nComputer use\nComputer use 主要操作数字界面，具身智能进一步进入物理世界。\n\nWorld model\n它常依赖世界模型来理解环境，并预测动作后果。\n\nAgent\nAgent 负责目标与规划，具身智能让它真正执行现实动作。", "relations": { "multimodal": { "label": "把感官接入…", "note": "它要同时处理视觉语音等多种输入。" }, "computer-use": { "label": "比…更进一步", "note": "前者操作电脑，后者直接碰现实世界。" }, "world-model": { "label": "依赖…理解环境", "note": "要行动稳，先得预测环境会怎么变。" }, "agent": { "label": "给…装上身体", "note": "让会规划的 AI 不只会说，还能动手。" } } } } }, { "id": "embodied-native-model", "name": "Embodied-native model", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "world-model" }, { "to": "robotics" } ], "track": "today", "seo": { "zh": { "title": "具身原生模型 是什么?驾校真上路,一文看懂 — AI Rookies", "description": "面向物理世界感知与行动训练的模型。人话版一分钟讲明白。" }, "en": { "title": "What Is Embodied-native model? Puppy Fetch Without Crashes", "description": "A model trained to sense the physical world and act inside it. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Embodied-native model", "factExplain": "A model trained to sense the physical world and act inside it.", "humanExplain": "An embodied-native model is like a puppy learning fetch in the yard. It chases the ball without crashing into the lawn chair.\n\nYou meet it in robots and self-driving cars. It helps AI handle space, movement, and feedback.", "humanExplainDisplay": "An embodied-native model is like\n==a puppy learning fetch== in the yard.\nIt chases the ball\nwithout ==crashing into the lawn chair==.\n\nYou meet it in robots and self-driving cars.\nIt helps AI handle space, movement, and feedback.", "relationsNarrative": "Embodied AI\nAn embodied-native model gives Embodied AI basic sensing and action skills.\n\nVLA\nVLA is a common way to connect vision, language, and action.\n\nWorld model\nA world model helps it predict what an action will cause.\n\nRobotics\nRobotics is the most direct place to give it a body.", "relations": { "embodied-ai": { "label": "serves as base for …", "note": "It is the model base for Embodied AI." }, "vision-language-action-model-vla": { "label": "can grow into …", "note": "VLA connects vision and language to action." }, "world-model": { "label": "internalizes …", "note": "A world model helps it predict physical results." }, "robotics": { "label": "lands in …", "note": "Robotics gives it the most direct body." } } }, "zh": { "fullName": "具身原生模型", "factExplain": "面向物理世界感知与行动训练的模型。", "humanExplain": "具身原生模型像驾校真上路：不只背交规，还会看镜、踩刹车、躲电动车。\n\n用于机器人、无人车和设备控制，让 AI 懂空间动作反馈。", "humanExplainDisplay": "具身原生模型像\n==驾校真上路==：\n不只背交规，\n还会看镜、踩刹车、躲电动车。\n\n用于机器人、无人车\n和设备控制，\n让 AI 懂空间动作反馈。", "relationsNarrative": "Embodied AI\n它为具身智能提供原生的感知与行动能力。\n\nVLA\nVLA 是它常见的视觉、语言到动作实现。\n\nWorld model\n世界模型让它能预判动作造成的后果。\n\nRobotics\n机器人是它最直接的落地载体。", "relations": { "embodied-ai": { "label": "作为…底座", "note": "它是具身智能里的模型底座。" }, "vision-language-action-model-vla": { "label": "发展成…", "note": "VLA 把视觉语言接到动作。" }, "world-model": { "label": "内化…", "note": "世界模型帮它预测物理后果。" }, "robotics": { "label": "落地到…", "note": "机器人是最典型的身体载体。" } } } } }, { "id": "embodied-video-model", "name": "Embodied Video Model", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "world-model" }, { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" } ], "track": "today", "i18n": { "en": { "fullName": "Embodied Video Model", "factExplain": "A model uses videos to learn how actions change the world.", "humanExplain": "Picture a robot eyeing a full cereal bowl. It plays the mess movie first: spoon, wobble, milk on socks.\n\nIt helps robots train and plan moves. The AI can guess what its actions will do.", "humanExplainDisplay": "Picture a robot eyeing a ==full cereal bowl==.\nIt plays the ==mess movie== first:\nspoon, wobble, milk on socks.\n\nIt helps robots train and plan moves.\nThe AI can guess\nwhat its actions will do.", "relationsNarrative": "World model\nAn Embodied Video Model is a video-based branch of a World model.\n\nEmbodied AI\nIt helps Embodied AI predict what real actions will cause.\n\nVLA\nA VLA can use it to preview a move before it acts.", "relations": { "world-model": { "label": "turns … into video", "note": "It predicts world changes as video after an action." }, "embodied-ai": { "label": "helps … act in the world", "note": "Embodied AI needs it to understand action results." }, "vision-language-action-model-vla": { "label": "helps … plan moves", "note": "A VLA can use it to test a move in video first." } } }, "zh": { "fullName": "具身视频模型", "factExplain": "用视频学习动作如何改变环境的模型。", "humanExplain": "具身视频模型像炒菜前脑内试锅：油会不会溅、菜会不会糊，想明白再下铲。\n\n用于机器人训练和动作规划，让AI预判现实因果。", "humanExplainDisplay": "具身视频模型像炒菜前\n==脑内试锅==：\n油会不会溅、菜会不会糊，\n==想明白再下铲==。\n\n用于机器人训练\n和动作规划，\n让AI预判现实因果。", "relationsNarrative": "World Model\n具身视频模型是世界模型的视频化分支。\n\nEmbodied AI\n它帮助具身 AI 预判真实行动的后果。\n\nVLA\nVLA 可借它先预演，再决定怎么动。", "relations": { "world-model": { "label": "把…视频化", "note": "它用视频预测动作后的世界变化。" }, "embodied-ai": { "label": "服务…落地", "note": "具身 AI 需要它理解行动后果。" }, "vision-language-action-model-vla": { "label": "辅助…规划动作", "note": "VLA 可用它预演动作是否可行。" } } } } }, { "id": "emergence", "name": "Emergence", "layer": "L1", "era": "2022", "publishedAt": "2026-05-23T11:45:00Z", "relations": [ { "to": "scaling-law" }, { "to": "foundation-model" }, { "to": "agi" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Emergent Abilities", "factExplain": "New abilities that suddenly appear when an AI model gets much bigger.", "humanExplain": "Emergence is like a kid learning words for months. One day they argue bedtime like a tiny lawyer.\n\nIt means AI skills do not always grow in a straight line. You see it in large LLMs and Foundation-models, so progress can surprise people.", "humanExplainDisplay": "Emergence is like a kid\n==learning words for months==.\nOne day they argue bedtime\nlike a ==tiny lawyer==.\n\nIt means AI skills do not always\ngrow in a straight line.\nYou see it in large LLMs\nand Foundation-models,\nso progress can surprise people.", "relationsNarrative": "Scaling-law\nScaling-law gives Emergence a size-growth background.\n\nFoundation-model\nBigger Foundation-models make Emergence easier to notice.\n\nAGI\nEmergence makes AGI seem like it could cross a sudden ability line.\n\nLLM\nLarge LLMs can show new Emergence abilities as they grow.", "relations": { "scaling-law": { "label": "gets context from …", "note": "Scaling-law gives Emergence a size-growth background." }, "foundation-model": { "label": "appears in large …", "note": "Bigger Foundation-models make Emergence easier to notice." }, "agi": { "label": "gets linked to …", "note": "Emergence makes AGI seem like it could cross a sudden ability line." }, "llm": { "label": "appears in large …", "note": "Large LLMs can show new Emergence abilities as they grow." } } }, "zh": { "fullName": "涌现能力", "factExplain": "模型规模提升后突然表现出的新能力。", "humanExplain": "涌现像孩子背了很久单词，某天突然能吵架了，家长既欣慰又害怕。\n\n它说明模型能力不总是线性增长，也让大模型发展带来更多不可预测性。", "humanExplainDisplay": "涌现像孩子背了很久单词，\n某天突然==会吵架了==。\n家长既欣慰，又有点害怕。\n\n它说明模型能力不总是线性增长。\n规模一大，\n惊喜和惊吓可能一起出现。", "relationsNarrative": "Scaling-law\nScaling-law 为 Emergence 提供规模增长的经验背景。\n\nFoundation-model\nFoundation-model 越大，越容易观察到 Emergence 现象。\n\nAGI\nEmergence 让 AGI 看起来可能跨过某个能力阈值。\n\nLLM\nLLM 规模持续扩大时，可能出现新的 Emergence 能力。", "relations": { "scaling-law": { "label": "由…解释" }, "foundation-model": { "label": "出现在大…中" }, "agi": { "label": "被联想到…" }, "llm": { "label": "出现在大…中" } } } } }, { "id": "empirical-risk-minimization", "name": "ERM", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "sgd" }, { "to": "bias-variance-tradeoff" }, { "to": "regularization" } ], "track": "history", "seo": { "en": { "title": "What Is Empirical Risk Minimization? The First Rule of Supervised Learning", "description": "ERM means making average training mistakes as small as possible — fit the data you have first, then worry about generalization. A plain-English explainer." }, "zh": { "title": "经验风险最小化是什么?监督学习的第一原则,一文看懂 — AI Rookies", "description": "先把训练集平均损失压到最低:眼前这批数据先拟合好,泛化后面再谈。ERM 是什么、边界在哪,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Empirical Risk Minimization", "factExplain": "Learning by making average training mistakes as small as possible.", "humanExplain": "ERM is like a kid studying only the practice test. They get those answers perfect, then hope the real test is not mean.\n\nIt is a basic goal in Supervised learning. The model first fits the training examples. Then we worry about new ones.", "humanExplainDisplay": "ERM is like a kid studying only the ==practice test==.\nThey get those answers ==perfect==,\nthen hope the real test is not mean.\n\nIt is a basic goal in Supervised learning.\nThe model first fits the training examples.\nThen we worry about new ones.", "relationsNarrative": "Supervised Learning\nERM is one common goal in Supervised learning.\n\nSGD\nSGD often lowers empirical risk one small step at a time.\n\nBias-Variance Tradeoff\nPushing empirical risk too low can hurt new answers.\n\nRegularization\nRegularization holds ERM back, so the model does not just memorize training examples.", "relations": { "supervised-learning": { "label": "sets a goal for …", "note": "Supervised learning often tries to lower training error." }, "sgd": { "label": "is optimized by …", "note": "SGD is often used to lower empirical risk step by step." }, "bias-variance-tradeoff": { "label": "is limited by …", "note": "Lower training error does not always mean better new answers." }, "regularization": { "label": "is restrained by …", "note": "Regularization stops the model from just memorizing the practice set." } } }, "zh": { "fullName": "经验风险最小化（Empirical Risk Minimization）", "factExplain": "通过最小化训练集平均损失来学习模型的基本原则。", "humanExplain": "网购店家先把详情页做到“看了都想下单”：眼前这批顾客先拿下，至于复购口碑，后面再说。\n\n它是监督学习的基本目标：先把训练数据拟合好，再考虑泛化。", "humanExplainDisplay": "网购店家先把详情页做到\n==“看了都想下单”==：\n眼前这批顾客先拿下，\n至于==复购口碑==，后面再说。\n\n它是监督学习的基本目标：\n先把训练数据拟合好，\n再考虑泛化。", "relationsNarrative": "Supervised-learning\n它是监督学习里最常见的优化目标之一。\n\nSGD\nSGD 常被用来一步步压低训练集上的经验风险。\n\nBias-variance-tradeoff\n经验风险降得太狠，可能换来更差的泛化表现。\n\nRegularization\n正则化会约束它，避免模型只会死记训练题。", "relations": { "supervised-learning": { "label": "构成…目标", "note": "监督学习常以最小化经验风险为目标。" }, "sgd": { "label": "靠…去优化", "note": "SGD 常用来实际最小化经验风险。" }, "bias-variance-tradeoff": { "label": "受…制约", "note": "一味压低训练误差未必泛化更好。" }, "regularization": { "label": "配合…防过拟合", "note": "正则化给只顾刷题的模型踩刹车。" } } } } }, { "id": "energy-based-model", "name": "Energy-Based Model", "layer": "L3", "era": "1980s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "boltzmann-machine" }, { "to": "markov-chain-monte-carlo" }, { "to": "probabilistic-graphical-model" } ], "track": "history", "i18n": { "en": { "fullName": "Energy-Based Model", "factExplain": "A model that scores how believable a sample is with an energy function.", "humanExplain": "An energy-based model gives every sample a weirdness score. Normal ones score low, strange ones score high. Walk the score downhill and it can even build or fix a sample.\n\nPeople use it to make examples, fix messy ones, or spot odd data. Training is slow because sampling is hard.", "humanExplainDisplay": "An energy-based model gives every sample\na ==weirdness score==.\nNormal ones score low,\nstrange ones score high.\n==Walk the score downhill==\nand it can build or fix a sample.\n\nPeople use it to make examples,\nfix messy ones, or spot odd data.\nTraining is slow\nbecause sampling is hard.", "relationsNarrative": "Generative Model\nAn Energy-Based Model can generate samples by looking for low-energy areas.\n\nBoltzmann Machine\nA Boltzmann Machine is a classic early Energy-Based Model.\n\nMCMC\nEnergy-Based Models often use MCMC to sample from low-energy areas.\n\nPGM\nMany Energy-Based Models can be seen as undirected PGMs.", "relations": { "generative-model": { "label": "belongs to … family", "note": "It can generate samples from low-energy areas." }, "boltzmann-machine": { "label": "continues … idea", "note": "A Boltzmann Machine is a classic Energy-Based Model." }, "markov-chain-monte-carlo": { "label": "often samples with …", "note": "MCMC helps find samples with low energy." }, "probabilistic-graphical-model": { "label": "can be written as …", "note": "Many Energy-Based Models can be seen as undirected PGMs." } } }, "zh": { "fullName": "能量模型", "factExplain": "用能量函数表示样本可信度的模型。", "humanExplain": "能量模型像给每张样本打“别扭分”：越自然分越低，越离谱分越高；顺着低分往下走，还能拼出或修好一张。\n\n能做生成、修复和挑异常，采样训练较费劲。", "humanExplainDisplay": "能量模型像给每张样本\n打==“别扭分”==：\n越自然分越低，\n越离谱分越高；\n==顺着低分往下走==，\n还能拼出或修好一张。\n\n能做生成、修复和挑异常，\n采样训练较费劲。", "relationsNarrative": "Generative Model\n它可通过寻找低能量区域来生成样本。\n\nBoltzmann Machine\n玻尔兹曼机是能量模型的经典早期形式。\n\nMCMC\n能量模型常用 MCMC 从低能量区域采样。\n\nProbabilistic Graphical Model\n许多能量模型可看作无向概率图模型。", "relations": { "generative-model": { "label": "属于…家族", "note": "它可从低能量区域生成样本。" }, "boltzmann-machine": { "label": "延续…思路", "note": "玻尔兹曼机是经典能量模型。" }, "markov-chain-monte-carlo": { "label": "常用…采样", "note": "低能量样本常靠 MCMC 找到。" }, "probabilistic-graphical-model": { "label": "可写成…", "note": "很多能量模型是无向概率图。" } } } } }, { "id": "ensemble-learning", "name": "Ensemble", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "gradient-boosting" }, { "to": "bias-variance-tradeoff" }, { "to": "classification" }, { "to": "regression" } ], "track": "history", "i18n": { "en": { "fullName": "Ensemble Learning", "factExplain": "A method that combines several models to get better results.", "humanExplain": "Ensemble Learning is like asking the whole lunch table for an answer. One kid may be way off, but the group is less silly.\n\nYou meet it in Classification. You meet it in Regression too. It makes results steadier, but it is harder to train and run.", "humanExplainDisplay": "Ensemble Learning is like asking\n==the whole lunch table== for an answer.\nOne kid may be way off,\nbut ==the group is less silly==.\n\nYou meet it in Classification.\nYou meet it in Regression too.\nIt makes results steadier,\nbut it is harder to train and run.", "relationsNarrative": "Gradient Boosting\nGradient Boosting is a classic path inside Ensemble Learning.\n\nBias-Variance Tradeoff\nIt often mixes several models to lower variance and make results steadier.\n\nClassification\nIn Classification, Ensemble Learning often uses voting to pick the final class.\n\nRegression\nIn Regression, Ensemble Learning often averages several predictions.", "relations": { "gradient-boosting": { "label": "includes …", "note": "Gradient Boosting is a major Ensemble Learning method." }, "bias-variance-tradeoff": { "label": "helps balance …", "note": "Ensembles often lower variance by mixing several models." }, "classification": { "label": "often used for …", "note": "In Classification, models can vote for the final class." }, "regression": { "label": "also used for …", "note": "In Regression, models can average their predictions." } } }, "zh": { "fullName": "集成学习（Ensemble Learning）", "factExplain": "把多个模型的结果组合起来提升整体表现的方法。", "humanExplain": "看中医时，老大夫不只摸一次脉就下结论，得望闻问切都对上，开的方子才不容易跑偏。\n\n常用于分类、回归等任务，提升稳定性，但训练部署更复杂。", "humanExplainDisplay": "看中医时，老大夫\n不只摸一次脉\n就下结论；得\n==望闻问切==都对上，\n开的方子才\n不容易==跑偏==。\n\n常用于分类、回归等任务，\n提升稳定性，\n但训练部署更复杂。", "relationsNarrative": "Gradient Boosting\n梯度提升是集成学习里很典型的一条技术路线。\n\nBias-Variance Tradeoff\n它常通过组合多个模型来降低方差、提升稳定性。\n\nClassification\n在分类任务里，它常用投票来决定最终类别。\n\nRegression\n在回归任务里，它常把多个预测做加权或平均。", "relations": { "gradient-boosting": { "label": "包含…这一路", "note": "梯度提升是集成学习的重要做法。" }, "bias-variance-tradeoff": { "label": "用来平衡…", "note": "常靠组合多个模型降低方差。" }, "classification": { "label": "常用于…任务", "note": "投票机制常见于分类场景。" }, "regression": { "label": "也用于…任务", "note": "可把多个预测结果做平均。" } } } } }, { "id": "enterprise-ai-deployment", "name": "Enterprise AI Deployment", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "llmops" }, { "to": "on-premise-ai" }, { "to": "api" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "Enterprise AI Deployment", "factExplain": "The process of putting an AI system into real company use.", "humanExplain": "It is like parking a taco truck inside the school cafeteria. Good tacos help. The principal still wants a permit.\n\nIt connects AI to real company work safely. You meet it in support desks, office tools, and knowledge bases.", "humanExplainDisplay": "It is like parking a ==taco truck==\ninside the school cafeteria.\nGood tacos help.\nThe ==principal still wants a permit==.\n\nIt connects AI to real company work safely.\nYou meet it in support desks,\noffice tools,\nand knowledge bases.", "relationsNarrative": "LLMOps\nEnterprise AI Deployment needs LLMOps for monitoring, testing, and rollback.\n\nOn-premise AI\nOn-premise AI is a common choice for rules and data boundaries.\n\nAPI\nAn API connects AI to current company systems and workflows.\n\nData-privacy\nEnterprise AI Deployment must handle access, storage, and leak risks first.", "relations": { "llmops": { "label": "goes live with …", "note": "LLMOps handles monitoring, testing, and rollback after launch." }, "on-premise-ai": { "label": "can use …", "note": "On-premise AI helps with rules and data boundaries." }, "api": { "label": "connects through …", "note": "An API lets company systems call the model." }, "data-privacy": { "label": "must protect …", "note": "Business data leaks are the big fear in company AI." } } }, "zh": { "fullName": "Enterprise AI Deployment（企业级 AI 部署）", "factExplain": "把 AI 系统落地到企业生产环境的过程。", "humanExplain": "企业级 AI 部署像路边摊开成中央厨房：偶尔惊艳不算数，得顿顿稳定、批量不翻车。\n\n让 AI 稳稳接业务，常见于客服、办公和知识库。", "humanExplainDisplay": "企业级 AI 部署像\n==路边摊开成中央厨房==：\n偶尔惊艳不算数，\n得==顿顿稳定、批量不翻车==。\n\n让 AI 稳稳接业务，\n常见于客服、办公和知识库。", "relationsNarrative": "LLMOps\n企业部署需要 LLMOps 做监控、评估和回滚。\n\nOn-premise AI\n本地部署是企业满足合规和数据边界的常见选择。\n\nAPI\nAPI 让企业把 AI 接进现有系统和流程。\n\nData-privacy\n企业部署必须先处理数据权限、留存和泄露风险。", "relations": { "llmops": { "label": "用…管上线", "note": "LLMOps 管上线后的监控、评估和回滚。" }, "on-premise-ai": { "label": "可选择…部署", "note": "本地部署常用于满足合规和数据边界。" }, "api": { "label": "常通过…接入", "note": "API 是企业系统调用模型的入口。" }, "data-privacy": { "label": "必须守住…", "note": "企业部署最怕业务数据外泄。" } } } } }, { "id": "ernie-bot", "name": "ERNIE Bot", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "chatbot" }, { "to": "llm" }, { "to": "chatgpt" } ], "track": "history", "i18n": { "en": { "fullName": "ERNIE Bot", "factExplain": "A Baidu generative AI chat product powered by the ERNIE large model.", "humanExplain": "ERNIE Bot is like a search box that took improv class. It can answer you, then help write the email.\n\nIt answers questions and helps with writing. It also helps with office work inside Baidu apps.", "humanExplainDisplay": "ERNIE Bot is like a ==search box==\nthat took ==improv class==.\nIt can answer you,\nthen help write the email.\n\nIt answers questions\nand helps with writing.\nIt also helps with office work\ninside Baidu apps.", "relationsNarrative": "Chatbot\nERNIE Bot is a chatbot product for talking with users.\n\nLLM\nERNIE Bot uses an LLM to understand questions and write answers.\n\nChatGPT\nERNIE Bot was an early China-market answer to ChatGPT.", "relations": { "chatbot": { "label": "is a type of …", "note": "It is a chat front door for users." }, "llm": { "label": "is powered by …", "note": "The LLM understands the question and writes the answer." }, "chatgpt": { "label": "is compared with …", "note": "It was an early China-market answer to ChatGPT." } } }, "zh": { "fullName": "文心一言", "factExplain": "基于文心大模型的生成式 AI 聊天产品。", "humanExplain": "文心一言像把百度搜索框练成相声搭子：不只报菜名，还能顺手写稿。\n\n它用于问答、写作和办公，也嵌入百度生态服务用户。", "humanExplainDisplay": "文心一言像把百度搜索框\n练成==相声搭子==：\n不只报菜名，\n还能顺手写稿。\n\n它用于问答、写作和办公，\n也嵌入百度生态，\n服务用户。", "relationsNarrative": "Chatbot\n文心一言是面向用户对话的聊天机器人产品。\n\nLLM\n它依赖大语言模型理解问题并生成回答。\n\nChatGPT\n它是中国市场早期对标 ChatGPT 的产品。", "relations": { "chatbot": { "label": "属于…", "note": "它本质是面向用户的聊天入口。" }, "llm": { "label": "由…驱动", "note": "背后模型负责理解并生成回答。" }, "chatgpt": { "label": "对标…", "note": "同属大众熟悉的聊天式 AI。" } } } } }, { "id": "expectation-maximization", "name": "EM", "layer": "L2", "era": "1977", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "unsupervised-learning" }, { "to": "kullback-leibler-divergence" }, { "to": "variational-inference" } ], "track": "history", "i18n": { "en": { "fullName": "Expectation-Maximization", "factExplain": "A method that takes turns guessing hidden data and updating model settings.", "humanExplain": "EM is like solving a mystery smoothie. You guess the hidden fruits, tweak the recipe, then guess again.\n\nIt helps when real labels are missing. You meet it in clustering and filling in missing data.", "humanExplainDisplay": "EM is like solving a ==mystery smoothie==.\nYou guess the ==hidden fruits==,\ntweak the recipe,\nthen guess again.\n\nIt helps when real labels are missing.\nYou meet it in clustering\nand filling in missing data.", "relationsNarrative": "Unsupervised Learning\nEM is often used to estimate hidden groups and settings without true labels.\n\nKL Divergence\nEM steps can be seen as slowly reducing a KL gap.\n\nVI\nVI can be seen as a wider version of EM.", "relations": { "unsupervised-learning": { "label": "is often used in …", "note": "EM estimates hidden structure when no true labels exist." }, "kullback-leibler-divergence": { "label": "can optimize …", "note": "Each EM round can shrink a KL gap." }, "variational-inference": { "label": "inspired …", "note": "VI is a broader version of EM." } } }, "zh": { "fullName": "Expectation-Maximization（期望最大化算法）", "factExplain": "一种在含隐变量时交替估计参数的优化方法。", "humanExplain": "像老中医号脉：先猜你偏哪种证候，再按这个猜法改方子；一轮轮试，病因画像就越来越准。\n\n常在没真标签时估参数，比如聚类或补全缺失信息。", "humanExplainDisplay": "像老中医号脉：\n先猜你偏哪种==证候==，\n再按这个猜法==改方子==；\n一轮轮试，病因画像就越来越准。\n\n常在没真标签时估参数，\n比如聚类或补全\n缺失信息。", "relationsNarrative": "Unsupervised Learning\n它常用于无监督场景下估计隐藏结构和参数。\n\nKL Divergence\n它的迭代过程常可写成逐步缩小 KL 差异。\n\nVariational Inference\n变分推断可看作把它推广到更一般的近似推断。", "relations": { "unsupervised-learning": { "label": "常用于…", "note": "常在没有真标签时迭代估计结构。" }, "kullback-leibler-divergence": { "label": "可视作优化…", "note": "其迭代过程常与 KL 目标相关。" }, "variational-inference": { "label": "启发了…", "note": "VI 可看作对它的推广版本。" } } } } }, { "id": "expert-system", "name": "Expert System", "layer": "L1", "era": "1970", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "symbolic-ai" }, { "to": "knowledge-representation" }, { "to": "production-system" } ], "track": "history", "i18n": { "en": { "fullName": "Expert System", "factExplain": "An AI system that uses expert rules and facts to make decisions.", "humanExplain": "An expert system is like a picky school nurse in a computer. If you have spots and a fever, it flips its rule cards.\n\nPeople used them for medical diagnosis and bank rules. They can explain why, but keeping the rules fresh is hard work.", "humanExplainDisplay": "An expert system is like a ==picky school nurse==\nin a computer.\nIf you have spots and a fever,\nit flips its ==rule cards==.\n\nPeople used them for medical diagnosis\nand bank rules.\nThey can explain why,\nbut keeping the rules fresh is hard work.", "relationsNarrative": "Symbolic AI\nExpert systems are a classic Symbolic AI way to solve problems with rules.\n\nKR\nExpert systems use KR to turn expert knowledge into machine-ready rules.\n\nProduction\nProduction often uses if-then rules to drive expert system reasoning.\n\nMYCIN\nMYCIN is a classic early medical expert system.", "relations": { "symbolic-ai": { "label": "belongs to …", "note": "Expert systems are a classic Symbolic AI example." }, "knowledge-representation": { "label": "depends on …", "note": "Rules and facts must be written for the machine to use." }, "production-system": { "label": "often runs on …", "note": "Production rules often drive the if-then reasoning." } } }, "zh": { "fullName": "专家系统", "factExplain": "用规则和知识库模拟专家决策的 AI 系统。", "humanExplain": "专家系统像修车老师傅进电脑：听异响看灯号，照“如果…就…”换件。\n\n曾用于诊断、金融规则，解释清楚，但维护很累。", "humanExplainDisplay": "专家系统像\n==修车老师傅==进电脑：\n听异响看灯号，\n照==“如果…就…”==换件。\n\n曾用于诊断、金融规则，\n解释清楚，\n但维护很累。", "relationsNarrative": "Symbolic AI\n专家系统是符号 AI 从规则出发解决问题的代表。\n\nKnowledge Representation\n专家系统依赖知识表示，把专家经验写成机器可用的形式。\n\nProduction\n产生式系统常用“如果…就…”规则驱动专家系统推理。\n\nMYCIN\nMYCIN 是早期医疗专家系统的经典案例。", "relations": { "symbolic-ai": { "label": "属于…路线", "note": "专家系统是符号 AI 的经典代表。" }, "knowledge-representation": { "label": "依赖…表达知识", "note": "规则和事实要先被机器看懂。" }, "production-system": { "label": "常用…执行规则", "note": "产生式规则驱动推理流程。" } } } } }, { "id": "explainable-ai", "name": "XAI", "layer": "L1", "era": "2016", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "reasoning-transparency" }, { "to": "model-uncertainty" }, { "to": "ai-bias" }, { "to": "ai-governance-framework" } ], "track": "history", "i18n": { "en": { "fullName": "Explainable AI", "factExplain": "Ways to make an AI model’s decisions easier for people to understand.", "humanExplain": "XAI is AI showing its work, like a kid in math class. No “trust me, I got 42” nonsense.\n\nYou meet it in hospitals. You meet it in banks. It helps review teams ask why the model decided that.", "humanExplainDisplay": "XAI is AI ==showing its work==,\nlike a kid in math class.\nNo “==trust me==, I got 42” nonsense.\n\nYou meet it in hospitals.\nYou meet it in banks.\nIt helps review teams ask\nwhy the model decided that.", "relationsNarrative": "Reasoning Transparency\nXAI asks if the model’s reasoning can be seen and explained.\n\nModel uncertainty\nModel uncertainty is the warning light XAI should show.\n\nAI-bias\nXAI can help people see where bias came from.\n\nAI Governance\nXAI lets regulators and auditors check more than the final result.", "relations": { "reasoning-transparency": { "label": "aims for …", "note": "Reasoning Transparency is the harder version of explainability." }, "model-uncertainty": { "label": "shows …", "note": "Model uncertainty says how unsure the model is." }, "ai-bias": { "label": "reveals …", "note": "Explanations can help show where bias came from." }, "ai-governance-framework": { "label": "supports …", "note": "AI Governance needs model explanations people can check." } } }, "zh": { "fullName": "可解释人工智能", "factExplain": "让模型决策过程更易被人理解的方法。", "humanExplain": "可解释AI要当老中医开方：不只写药名，还得讲哪味治哪症、凭啥这么抓。\n\n它用于医疗、金融、审核，帮助追问模型判断依据。", "humanExplainDisplay": "可解释AI要当\n==老中医开方==：\n不只写药名，\n还得讲==哪味治哪症==、凭啥这么抓。\n\n它用于医疗、金融、审核，\n帮助追问模型判断依据。", "relationsNarrative": "Reasoning Transparency\n它关注模型推理过程能否被看见、讲清。\n\nModel Uncertainty\n不确定性是解释里最该亮出的警示灯。\n\nAI Bias\n解释能帮助人发现偏见来自哪里。\n\nAI Governance\n可解释性让监管和审计不只看结果。", "relations": { "reasoning-transparency": { "label": "追求…", "note": "透明推理是可解释性的高难版本。" }, "model-uncertainty": { "label": "亮出…", "note": "不确定性说明模型有多没把握。" }, "ai-bias": { "label": "暴露…", "note": "解释能帮助发现偏见从哪来。" }, "ai-governance-framework": { "label": "支撑…", "note": "治理需要能追责的模型说明。" } } } } }, { "id": "exploration-exploitation", "name": "Exploration-Exploitation Tradeoff", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "q-learning" }, { "to": "temperature" }, { "to": "policy-gradient" } ], "track": "history", "i18n": { "en": { "fullName": "Exploration-Exploitation Tradeoff", "factExplain": "Balancing trying new choices with using choices that already work.", "humanExplain": "It is the ice-cream problem. Do you order trusty chocolate, or risk pickle swirl?\n\nYou meet it in RL, recommendation feeds, and online ads. It decides if a system sticks with a winner or tries a new path.", "humanExplainDisplay": "It is the ==ice-cream problem==.\nDo you order ==trusty chocolate==,\nor risk pickle swirl?\n\nYou meet it in RL,\nrecommendation feeds,\nand online ads.\nIt decides if a system sticks with a winner\nor tries a new path.", "relationsNarrative": "RL\nThis is one of RL's classic choices.\n\nQ-Learning\nQ-Learning balances top-scoring actions with new actions to test.\n\nTemperature\nTemperature changes randomness, so it changes how much the system explores.\n\nPolicy Gradient\nPolicy Gradient must chase steady rewards and keep room for trial and error.", "relations": { "reinforcement-learning": { "label": "is a core problem in …", "note": "RL keeps facing this choice." }, "q-learning": { "label": "guides action choice in …", "note": "A high Q value is not always the move to pick." }, "temperature": { "label": "can be tuned by …", "note": "Higher Temperature makes the system try less obvious choices." }, "policy-gradient": { "label": "shapes updates in …", "note": "Policy Gradient needs steady wins and some room to try." } } }, "zh": { "fullName": "探索-利用权衡", "factExplain": "在尝试新选择与利用已知好选择间做平衡。", "humanExplain": "像租房时老盯着那套最顺眼的最省心，可你不多看几家，就不知道下一套会不会更值。\n\n常见于强化学习、推荐和投放，影响系统敢不敢试新路。", "humanExplainDisplay": "像租房时老盯着那套==最顺眼==的最省心，\n可你不多看几家，\n就不知道下一套会不会更==值==。\n\n常见于强化学习、\n推荐和投放，\n影响系统敢不敢试新路。", "relationsNarrative": "Reinforcement-learning\n它是强化学习里最经典的决策权衡之一。\n\nQ-Learning\nQ-Learning 需要在高分动作和试新动作间平衡。\n\nTemperature\n温度可影响采样随机性，从而改变探索程度。\n\nPolicy Gradient\n策略优化时，也要兼顾稳定收益和试错空间。", "relations": { "reinforcement-learning": { "label": "是…核心难题", "note": "强化学习几乎绕不开这道选择题。" }, "q-learning": { "label": "决定…选动作", "note": "Q 值高也未必次次都该照着选。" }, "temperature": { "label": "可用…调探索", "note": "温度高一点，选项会更敢试新路。" }, "policy-gradient": { "label": "影响…策略更新", "note": "策略既要求稳，也得留点试错空间。" } } } } }, { "id": "face-recognition", "name": "Face Recognition", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "cnn" }, { "to": "ai-bias" }, { "to": "data-privacy" } ], "track": "history", "i18n": { "en": { "fullName": "Face Recognition", "factExplain": "A technology for identifying or checking a person’s face in an image.", "humanExplain": "Face recognition is the bouncer at your phone’s tiny nightclub. Show your face, and the velvet rope opens.\n\nYou meet it when phones unlock or office doors open. Security cameras use it too, but privacy and unfair errors are real risks.", "humanExplainDisplay": "Face recognition is the ==bouncer==\nat your phone’s tiny nightclub.\nShow your face,\nand the ==velvet rope opens==.\n\nYou meet it when phones unlock\nor office doors open.\nSecurity cameras use it too,\nbut privacy and unfair errors are real risks.", "relationsNarrative": "Computer Vision\nFace recognition is a classic identity task in Computer Vision.\n\nCNN\nCNNs can pull face features from images for matching.\n\nAI-bias\nUneven data can make error rates higher for some groups.\n\nData-privacy\nFaces are sensitive body data, so collection and storage need care.", "relations": { "computer-vision": { "label": "is a … task", "note": "Face recognition is a classic identity task in Computer Vision." }, "cnn": { "label": "often used … to find features", "note": "CNNs were widely used to pull useful face features from images." }, "ai-bias": { "label": "can amplify …", "note": "Uneven training data can cause unfair face recognition errors." }, "data-privacy": { "label": "touches …", "note": "Face data is highly sensitive personal data." } } }, "zh": { "fullName": "人脸识别", "factExplain": "识别或验证图像中人脸身份的技术。", "humanExplain": "人脸识别像小区门口保安：你一摘口罩，它就翻脑内住户册。\n\n用于解锁、门禁和安防，也带来隐私与偏见风险。", "humanExplainDisplay": "人脸识别像小区门口\n==保安==：\n你一摘口罩，\n它就翻脑内==住户册==。\n\n用于解锁、门禁和安防，\n也带来隐私与偏见风险。", "relationsNarrative": "Computer Vision\n人脸识别是计算机视觉里的典型身份识别任务。\n\nCNN\nCNN 常用于从人脸图像中提取可比对特征。\n\nAI-bias\n数据分布不均，可能让不同人群误识率不同。\n\nData-privacy\n人脸属于敏感生物信息，采集和存储都需谨慎。", "relations": { "computer-vision": { "label": "属于…任务", "note": "人脸识别是典型视觉识别应用。" }, "cnn": { "label": "常用…提特征", "note": "CNN 曾是人脸特征提取主力。" }, "ai-bias": { "label": "可能放大…", "note": "训练数据不均会带来识别偏差。" }, "data-privacy": { "label": "触及…", "note": "人脸数据天然高度敏感。" } } } } }, { "id": "factor-graph", "name": "Factor Graph", "layer": "L2", "era": "1990s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "probabilistic-graphical-model" }, { "to": "belief-propagation" }, { "to": "bayesian-network" }, { "to": "slam" } ], "track": "history", "i18n": { "en": { "fullName": "Factor Graph", "factExplain": "A graph that breaks a big probability model into variables and small factor pieces.", "humanExplain": "A factor graph is like small group chats for one big family. Each chat holds one matter and only the people it touches. News hops from chat to chat until everyone agrees.\n\nYou meet it in probability inference and SLAM. It draws who depends on whom, then passes messages along the links.", "humanExplainDisplay": "A factor graph is like\n==small group chats==\nfor one big family.\nEach chat holds one matter\nand only the people it touches.\nNews ==hops from chat to chat==\nuntil everyone agrees.\n\nYou meet it in probability inference and SLAM.\nIt draws who depends on whom,\nthen passes messages along the links.", "relationsNarrative": "PGM\nA factor graph is a variable-and-factor form of a PGM.\n\nBP\nBP often sends local messages back and forth on a factor graph.\n\nBayesian Network\nA Bayesian Network can become a factor graph for easier inference.\n\nSLAM\nSLAM often uses factor graphs to combine sensor clues.", "relations": { "probabilistic-graphical-model": { "label": "breaks down …", "note": "A factor graph is a split-up form of a PGM." }, "belief-propagation": { "label": "carries …", "note": "BP often passes small local messages on a factor graph." }, "bayesian-network": { "label": "rewrites …", "note": "A Bayesian Network can be turned into a factor graph for inference." }, "slam": { "label": "supports …", "note": "SLAM often uses factor graphs to combine sensor clues." } } }, "zh": { "fullName": "因子图", "factExplain": "用变量和因子表示联合分布分解的图模型。", "humanExplain": "因子图像把大家族拉成小群：一件事一个群，谁跟谁有关一眼看清，消息只在群里传。\n\n用于概率推断和 SLAM，把依赖画成关系网分头算。", "humanExplainDisplay": "因子图像把大家族\n拉成小群：\n一件事==一个群==，\n谁跟谁有关一眼看清，\n消息只在==群里传==。\n\n用于概率推断和 SLAM，\n把依赖画成关系网，\n分头算。", "relationsNarrative": "PGM\n因子图是 PGM 的一种变量—因子分解表示。\n\nBP\nBP 常在因子图上，把局部消息来回传递。\n\nBayesian Network\n贝叶斯网络可改写成因子图，方便推断。\n\nSLAM\nSLAM 常用因子图融合传感器约束。", "relations": { "probabilistic-graphical-model": { "label": "细化…", "note": "因子图是 PGM 的分解表示。" }, "belief-propagation": { "label": "承载…", "note": "BP 常在因子图上传递局部消息。" }, "bayesian-network": { "label": "改写…", "note": "贝叶斯网络可转成因子图来推断。" }, "slam": { "label": "支撑…", "note": "SLAM 常用因子图融合传感器约束。" } } } } }, { "id": "fallback-model", "name": "Fallback model", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "model-routing" }, { "to": "llmops" }, { "to": "api" }, { "to": "cost-aware-ai-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Fallback Model", "factExplain": "A backup model that handles requests when the main model is down.", "humanExplain": "A fallback model is the spare cashier who opens lane 3. The main line jams, and nobody has to glare at the candy rack.\n\nIt steps in during rush hours, outages, or usage limits. The app keeps answering instead of going dark.", "humanExplainDisplay": "A fallback model is the ==spare cashier==\nwho opens lane 3.\nThe main line jams,\nand ==nobody has to glare==\nat the candy rack.\n\nIt steps in during rush hours,\noutages,\nor usage limits.\nThe app keeps answering\ninstead of going dark.", "relationsNarrative": "Routing\nRouting decides when a request moves to the fallback model.\n\nLLMOps\nLLMOps uses a fallback model to keep AI services running.\n\nAPI\nAn API can hide the switch behind one call.\n\nCost-aware AI\nCost-aware AI can use a cheaper fallback model during busy times.", "relations": { "model-routing": { "label": "switches by …", "note": "Routing decides when to use the fallback model." }, "llmops": { "label": "supports … reliability", "note": "A fallback model helps keep live AI services running." }, "api": { "label": "connects through …", "note": "An API often hides the main-to-fallback switch." }, "cost-aware-ai-ai": { "label": "helps … control cost", "note": "A cheaper fallback model can cover busy times." } } }, "zh": { "fullName": "备用模型", "factExplain": "主模型不可用时接管请求的备用模型。", "humanExplain": "备用模型就是地铁备用线：主线一堵，先把人绕出去，别让服务趴站台。\n\n用于高峰、故障、限额时，兜住请求不中断。", "humanExplainDisplay": "备用模型就是\n==地铁备用线==：\n主线一堵，先把人绕出去，\n别让服务==趴站台==。\n\n用于高峰、故障、限额时，\n兜住请求不中断。", "relationsNarrative": "Model-routing\n路由决定请求何时从主模型切到备用模型。\n\nLLMOps\n备用模型是线上 AI 服务的可靠性兜底。\n\nAPI\nAPI 常把主备模型切换封装在统一接口后。\n\nCost-aware AI\n备用模型也可在高峰时用来控制成本。", "relations": { "model-routing": { "label": "依赖…切换", "note": "路由决定何时换到备用模型。" }, "llmops": { "label": "服务于…可靠性", "note": "备用模型是线上运维的兜底手段。" }, "api": { "label": "通过…接入", "note": "API 常把主备模型切换藏在接口后。" }, "cost-aware-ai-ai": { "label": "配合…控成本", "note": "高峰时可用便宜模型先兜底。" } } } } }, { "id": "faster-r-cnn", "name": "Faster R-CNN", "layer": "L3", "era": "2015", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "r-cnn" }, { "to": "object-detection" }, { "to": "cnn" }, { "to": "yolo" } ], "track": "history", "i18n": { "en": { "fullName": "Faster Region-based Convolutional Neural Network", "factExplain": "An object detection model with a region proposal network for finding objects.", "humanExplain": "Faster R-CNN is like a school lunch monitor. It first circles the flying-fries table. Then it checks who caused the mess.\n\nIt first draws likely boxes in an image. Then it names what is inside. It can find people and cars. It can also mark spots in medical scans. It is steady and accurate. But it is slower than one-stage models.", "humanExplainDisplay": "Faster R-CNN is like a ==school lunch monitor==.\nIt first circles the ==flying-fries table==.\nThen it checks who caused the mess.\n\nIt first draws likely boxes in an image.\nThen it names what is inside.\nIt can find people and cars.\nIt can also mark spots in medical scans.\nIt is steady and accurate.\nBut it is slower than one-stage models.", "relationsNarrative": "R-CNN\nFaster R-CNN moves region proposal generation into the network.\n\nObject Detection\nFaster R-CNN is a classic two-stage model for object detection.\n\nCNN\nFaster R-CNN uses a CNN to pull useful features from images.\n\nYOLO\nYOLO detects in one step, so it leans harder toward speed.", "relations": { "r-cnn": { "label": "improves on …", "note": "It moves region proposals into the network." }, "object-detection": { "label": "is used for …", "note": "It outputs object classes and boxes." }, "cnn": { "label": "uses … for features", "note": "CNN features help the model recognize image regions." }, "yolo": { "label": "contrasts with …", "note": "YOLO detects in one step, so it is usually faster." } } }, "zh": { "fullName": "快速区域卷积神经网络", "factExplain": "一种用区域提议网络做目标检测的模型。", "humanExplain": "Faster R-CNN像武馆师傅挑对手：先圈出有招的人，再细看是哪门派。\n\n用于行人、车辆、病灶检测，精度稳，速度逊于一阶段。", "humanExplainDisplay": "Faster R-CNN像\n武馆师傅挑对手：\n==先圈出有招的人==，\n再细看是哪门派。\n\n用于行人、车辆、病灶检测，\n精度稳，\n速度逊于一阶段。", "relationsNarrative": "R-CNN\nFaster R-CNN 把候选区域生成并入网络。\n\nObject Detection\n它是目标检测的经典两阶段模型。\n\nCNN\n它借助卷积网络提取图像特征。\n\nYOLO\nYOLO 一步完成检测，速度取向更明显。", "relations": { "r-cnn": { "label": "改进…", "note": "它把候选区域生成并入网络。" }, "object-detection": { "label": "用于…", "note": "它输出物体类别和位置框。" }, "cnn": { "label": "借助…提特征", "note": "卷积特征让图像区域可被识别。" }, "yolo": { "label": "对比…", "note": "YOLO 一步检测，更快但取舍不同。" } } } } }, { "id": "fasttext", "name": "FastText", "layer": "L3", "era": "2016", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "word2vec" }, { "to": "embedding" }, { "to": "text-classification" }, { "to": "distributional-semantics" } ], "track": "history", "i18n": { "en": { "fullName": "Subword-Enhanced Text Representation and Classification", "factExplain": "A method that learns word vectors from word parts and classifies text fast.", "humanExplain": "FastText is like a kid reading a giant cereal-box word. It spots small pieces it knows, then makes a smart guess.\n\nIt learns light word embeddings from word pieces. It also sorts text fast, like news stories or reviews.", "humanExplainDisplay": "FastText is like a kid reading\n==a giant cereal-box word==.\nIt spots ==small pieces it knows==,\nthen makes a smart guess.\n\nIt learns light word embeddings\nfrom word pieces.\nIt also sorts text fast,\nlike news stories or reviews.", "relationsNarrative": "Word2Vec\nFastText builds on Word2Vec by adding word-piece information.\n\nEmbedding\nFastText is often used to train small, reusable word embeddings.\n\nText Cls\nFastText can quickly train classifiers for news, reviews, and other text.\n\nDist. Semantics\nFastText still learns word meaning from nearby words.", "relations": { "word2vec": { "label": "extends …", "note": "It adds word pieces to the Word2Vec idea." }, "embedding": { "label": "produces …", "note": "It is often used to train small word embeddings." }, "text-classification": { "label": "powers …", "note": "It can train a text classifier very quickly." }, "distributional-semantics": { "label": "inherits …", "note": "It still learns meaning from words seen nearby." } } }, "zh": { "fullName": "子词增强文本表示与分类方法", "factExplain": "用子词信息学习词向量并快速做文本分类的方法。", "humanExplain": "FastText 像修车师傅认零件：整车没见过，螺丝链条熟，也能猜个八九不离十。\n\n用于轻量词向量和文本分类，低成本好上手。", "humanExplainDisplay": "FastText 像修车师傅认零件：\n整车没见过，\n==螺丝链条熟==，\n也能猜个==八九不离十==。\n\n用于轻量词向量和文本分类，\n低成本好上手。", "relationsNarrative": "Word2Vec\nFastText 在 Word2Vec 基础上加入子词信息。\n\nEmbedding\n它常用于训练轻量、可迁移的词向量。\n\nText Classification\n它能快速训练新闻、评论等文本分类器。\n\nDistributional Semantics\n它仍依赖上下文共现来学习词义。", "relations": { "word2vec": { "label": "扩展…的思路", "note": "它把词拆成子词来学向量。" }, "embedding": { "label": "产出…", "note": "它常用来训练轻量词向量。" }, "text-classification": { "label": "服务…", "note": "它也能快速训练文本分类器。" }, "distributional-semantics": { "label": "继承…", "note": "词义仍来自上下文共现规律。" } } } } }, { "id": "feature-engineering", "name": "Feature-engineering", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "representation-learning" }, { "to": "supervised-learning" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Feature Engineering", "factExplain": "Turning raw data into inputs a model can use better.", "humanExplain": "Feature engineering is taco night prep. Chop the toppings and label the lava salsa before Uncle Dave grabs it.\n\nYou meet it in spreadsheet data. For old-school models, it can make or break risk checks and recommendations.", "humanExplainDisplay": "Feature engineering is ==taco night prep==.\nChop the toppings\nand ==label the lava salsa==\nbefore Uncle Dave grabs it.\n\nYou meet it in spreadsheet data.\nFor old-school models,\nit can make or break\nrisk checks and recommendations.", "relationsNarrative": "Representation Learning\nRepresentation Learning lets the model learn features itself, so people hand-build fewer.\n\nSupervised Learning\nFeature engineering often prepares the input before Supervised Learning trains.\n\nDeep Learning\nDeep Learning made many hand-built features less important.", "relations": { "representation-learning": { "label": "often replaced by …", "note": "Representation Learning lets the model learn features itself." }, "supervised-learning": { "label": "prepares data for …", "note": "Feature engineering shapes inputs before Supervised Learning trains." }, "deep-learning": { "label": "matters less in …", "note": "Deep Learning cuts down much hand-built feature work." } } }, "zh": { "fullName": "Feature Engineering／特征工程", "factExplain": "把原始数据加工成更适合模型使用的输入表示。", "humanExplain": "特征工程像夜市摆摊备料：串怎么签、辣度怎么分、招牌怎么摆，收拾顺了，客人下单才更痛快。\n\n常见于表格数据、风控和推荐，直接影响传统模型效果。", "humanExplainDisplay": "特征工程像夜市摆摊\n备==料==：\n串怎么签、辣度怎么分，\n招牌怎么摆，\n收拾顺了，客人下单\n才更==痛快==。\n\n常见于表格数据、\n风控和推荐，\n直接影响传统模型效果。", "relationsNarrative": "Representation Learning\n表示学习让模型自己学特征，减少手工设计。\n\nSupervised Learning\n它常作为监督学习前的数据处理步骤。\n\nDeep Learning\n深度学习兴起后，手工特征的重要性下降。", "relations": { "representation-learning": { "label": "常被…替代", "note": "深度学习常自动学表示，少手工做特征。" }, "supervised-learning": { "label": "作为…前处理", "note": "先把输入整理好，再交给监督学习建模。" }, "deep-learning": { "label": "在…中弱化", "note": "深度模型减少了大量手工特征设计。" } } } } }, { "id": "feature-selection", "name": "Feature Selection", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "feature-engineering" }, { "to": "lasso" }, { "to": "regularization" }, { "to": "decision-tree" } ], "track": "history", "i18n": { "en": { "fullName": "Feature Selection", "factExplain": "A way to keep the most useful inputs and drop the rest.", "humanExplain": "Feature selection is like cleaning out a kitchen junk drawer. Keep the batteries. Toss the mystery keys.\n\nYou use it with table data. It cuts noise. It makes models faster. It makes results easier to explain.", "humanExplainDisplay": "Feature selection is like cleaning out\n==a kitchen junk drawer==.\nKeep the ==batteries==.\nToss the mystery keys.\n\nYou use it with table data.\nIt cuts noise.\nIt makes models faster.\nIt makes results easier to explain.", "relationsNarrative": "Feature-engineering\nFeature selection is a key step in feature engineering.\n\nLasso\nLasso can push weak feature weights to zero.\n\nRegularization\nFeature selection often works with regularization to reduce overfitting.\n\nDecision Tree\nA decision tree naturally favors more useful features when it splits.", "relations": { "feature-engineering": { "label": "is part of …", "note": "Feature engineering makes inputs. Feature selection keeps the useful ones." }, "lasso": { "label": "can use …", "note": "Lasso can push some feature weights down to zero." }, "regularization": { "label": "controls complexity with …", "note": "Dropping weak features can help reduce overfitting." }, "decision-tree": { "label": "is often done by …", "note": "A tree naturally favors stronger features when it splits." } } }, "zh": { "fullName": "特征选择", "factExplain": "从已有特征中挑出更有用子集的方法。", "humanExplain": "它像相亲简历去水分：头衔再花也先划掉，最后留下真能说明人的那几条。\n\n常用于表格数据建模，帮模型降噪提速，也让结果更容易解释。", "humanExplainDisplay": "它像相亲简历==去水分==：\n头衔再花也==先划掉==，\n最后留下真能说明人的\n那几条。\n\n常用于表格数据建模，\n帮模型降噪提速，\n也让结果更容易解释。", "relationsNarrative": "Feature-engineering\n特征选择是特征工程里的关键一步，负责留下有用输入。\n\nLasso\nLasso 会把部分权重压成零，常被用来做特征筛选。\n\nRegularization\n它常和正则化一起用，减少过拟合与无效信息。\n\nDecision Tree\n决策树在分裂过程中，也会自然偏向更有用的特征。", "relations": { "feature-engineering": { "label": "属于…的一环", "note": "先造特征，再挑该留下哪些。" }, "lasso": { "label": "可用…来做", "note": "L1 正则会把部分特征权重压到零。" }, "regularization": { "label": "常借…控复杂度", "note": "减少无用特征，能缓解过拟合。" }, "decision-tree": { "label": "常被…隐式完成", "note": "树模型分裂时会自然筛掉弱特征。" } } } } }, { "id": "federated-learning", "name": "FL", "layer": "L2", "era": "2010s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "pretraining" }, { "to": "on-premise-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Federated Learning", "factExplain": "A way to train one model by sharing updates, not local data.", "humanExplain": "Federated learning is a group project with locked backpacks. Everyone shares what they learned, not the private homework.\n\nYou meet it in phone keyboards. It also helps hospitals and banks train together without pooling raw data.", "humanExplainDisplay": "Federated learning is a group project\nwith ==locked backpacks==.\nEveryone shares ==what they learned==,\nnot the private homework.\n\nYou meet it in phone keyboards.\nIt also helps hospitals and banks\ntrain together without pooling raw data.", "relationsNarrative": "Data-privacy\nFederated learning lowers privacy pressure by keeping data local.\n\nPretraining\nFederated learning can train models without pooling raw data.\n\nOn-premise AI\nFederated learning is useful when data must stay on local machines.", "relations": { "data-privacy": { "label": "helps protect …", "note": "One main goal is to touch raw data as little as possible." }, "pretraining": { "label": "can be used in …", "note": "It is one way to organize distributed training." }, "on-premise-ai": { "label": "fits …", "note": "Keeping data local often matches on-premise needs." } } }, "zh": { "fullName": "Federated learning｜联邦学习", "factExplain": "一种数据不出本地、只汇总更新来训练模型的方法。", "humanExplain": "各家数据锁死在自己地盘、绝不外传，只派代表出来对个学习心得，活像武林各派切磋：招式互通，秘籍不借。\n\n适合输入法、医疗、金融等敏感场景，不汇总原始数据也能联手训练。", "humanExplainDisplay": "各家数据锁死在自己地盘、绝不外传，\n只派代表出来==对个学习心得==，\n活像==武林各派切磋==：\n招式互通，秘籍不借。\n\n适合输入法、医疗、金融等敏感场景，\n不汇总原始数据也能联手训练。", "relationsNarrative": "Data-privacy\n它通过数据留本地，来降低集中收集的隐私压力。\n\nPretraining\n它可用于模型训练阶段，只是不集中原始数据。\n\nOn-premise AI\n当数据不能离开本地时，这种方式尤其合适。", "relations": { "data-privacy": { "label": "兼顾…需求", "note": "核心目标之一就是少碰原始数据。" }, "pretraining": { "label": "可用于…阶段", "note": "它是分布式训练的一种组织方式。" }, "on-premise-ai": { "label": "适合…场景", "note": "数据留在本地更符合这类部署诉求。" } } } } }, { "id": "few-shot-learning", "name": "Few-Shot Learning", "layer": "L2", "era": "2000s", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "in-context-learning" }, { "to": "transfer-learning" }, { "to": "fine-tuning" } ], "track": "history", "i18n": { "en": { "fullName": "Few-Shot Learning", "factExplain": "A way for AI to learn a new task from just a few examples.", "humanExplain": "Few-shot learning is like peeking at two solved math problems before a quiz. Somehow, you fake the next one well enough to pass.\n\nIt helps when labeled examples are rare. The AI learns a new task from just a few samples.", "humanExplainDisplay": "Few-shot learning is like\n==peeking at two solved math problems==\nbefore a quiz.\nSomehow, you ==fake the next one==\nwell enough to pass.\n\nIt helps when labeled examples are rare.\nThe AI learns a new task\nfrom just a few samples.", "relationsNarrative": "In-Context Learning\nFew-shot learning often comes from examples inside the prompt.\n\nTransfer Learning\nIt builds on general skills learned before.\n\nFine-tuning\nIt can also use a small labeled set to tune the model.", "relations": { "in-context-learning": { "label": "often uses …", "note": "Examples in the prompt teach the model the task for now." }, "transfer-learning": { "label": "builds on …", "note": "General skills move to a new task with only a few examples." }, "fine-tuning": { "label": "can use …", "note": "A small labeled set can gently tune the model." } } }, "zh": { "fullName": "Few-Shot Learning｜少样本学习", "factExplain": "只用少量样本就学会新任务的学习方式。", "humanExplain": "少样本学习像考前抱佛脚翻两道例题，真上考场居然也能照猫画虎写个差不多。\n\n适合标注稀缺的新任务，用少量样本也能快速上手。", "humanExplainDisplay": "少样本学习像考前抱佛脚翻两道例题，\n真上考场居然也能==照猫画虎==写个==差不多==。\n\n适合标注稀缺的新任务，\n用少量样本也能快速上手。", "relationsNarrative": "In-Context Learning\n大模型里的少样本能力，常靠提示中的示例触发。\n\nTransfer Learning\n它通常建立在已有通用能力迁移到新任务上。\n\nFine-tuning\n少样本学习也可通过少量数据微调来实现。", "relations": { "in-context-learning": { "label": "常靠…实现", "note": "大模型常用示例提示临时学会任务。" }, "transfer-learning": { "label": "建立在…上", "note": "先学通用能力，再少量样本适配。" }, "fine-tuning": { "label": "可用…落地", "note": "少量标注数据也能做轻量适配。" } } } } }, { "id": "fine-tuning", "name": "Fine-tuning", "layer": "L2", "era": "2018", "publishedAt": "2026-05-23T08:30:00Z", "relations": [ { "to": "pretraining" }, { "to": "foundation-model" }, { "to": "rlhf" }, { "to": "distillation" } ], "i18n": { "en": { "fullName": "Fine-tuning", "factExplain": "Training an existing model more with specific data to fit a task.", "humanExplain": "Fine-tuning is boot camp for a smart new hire. They come back knowing your weird forms and Slack slang.\n\nIt helps with repeat work in one style or field. It is not a magic wish fountain.", "humanExplainDisplay": "Fine-tuning is ==boot camp==\nfor a smart new hire.\nThey come back knowing\n==your weird forms and Slack slang==.\n\nIt helps with repeat work\nin one style or field.\nIt is not a magic wish fountain.", "relationsNarrative": "Pretraining\nFine-tuning comes after Pretraining and keeps adjusting the model.\n\nFoundation-model\nFine-tuning helps a Foundation-model fit a specific task.\n\nRLHF\nRLHF is often used during Fine-tuning to teach preferences.\n\nDistillation\nDistillation can move Fine-tuning gains into a smaller model.", "relations": { "pretraining": { "label": "comes after …", "note": "Fine-tuning starts after Pretraining and adjusts the model further." }, "foundation-model": { "label": "adapts … to a task", "note": "Fine-tuning makes a Foundation-model fit one job better." }, "rlhf": { "label": "can include …", "note": "RLHF is often a fine-tuning step for teaching preferences." }, "distillation": { "label": "often pairs with …", "note": "Distillation can move fine-tuned skills into a smaller model." } } }, "zh": { "fullName": "微调", "factExplain": "在已有模型基础上用特定数据继续训练以适配任务。", "humanExplain": "微调像给全能打工人报个岗前培训，不重读大学，只学公司黑话。\n\n它让模型更贴合特定风格、任务或行业，常用于客服、写作和垂直助手。", "humanExplainDisplay": "微调像给==全能打工人==\n报个==岗前培训==，\n不重读大学，\n只学公司黑话。\n\n它让模型更贴合\n特定风格、任务或行业，\n常用于客服、写作\n和垂直助手。", "relationsNarrative": "Pretraining\nFine-tuning 接在 Pretraining 之后，继续调整模型能力。\n\nFoundation-model\nFoundation-model 经过 Fine-tuning 后更适配具体任务。\n\nRLHF\nRLHF 常作为 Fine-tuning 的一类偏好优化步骤。\n\nDistillation\nDistillation 可把 Fine-tuning 后的能力迁移到小模型。", "relations": { "pretraining": { "label": "在…之后进行" }, "foundation-model": { "label": "让…适配任务" }, "rlhf": { "label": "可结合…" }, "distillation": { "label": "常配合…" } } } } }, { "id": "flash-attention", "name": "Flash Attention", "layer": "L5", "sublayer": "compute", "era": "2022", "publishedAt": "2026-05-30T03:10:23.230Z", "relations": [ { "to": "attention" }, { "to": "transformer" }, { "to": "vram" }, { "to": "llm-inference-engine" } ], "track": "today", "i18n": { "en": { "fullName": "Flash Attention", "factExplain": "A faster Attention method with lower VRAM use.", "humanExplain": "Flash Attention is like washing dishes while you cook. The sink stays empty, and dinner lands sooner.\n\nIt speeds up Attention and uses less VRAM. You meet it during training and inference. It helps with long text.", "humanExplainDisplay": "Flash Attention is like ==washing dishes while you cook==.\nThe sink stays ==empty==,\nand dinner lands sooner.\n\nIt speeds up Attention\nand uses less VRAM.\nYou meet it during training and inference.\nIt helps with long text.", "relationsNarrative": "Attention\nFlash Attention makes Attention faster and uses less VRAM.\n\nTransformer\nTransformers rely on Attention, so Flash Attention often speeds them up.\n\nVRAM\nFlash Attention uses less temporary memory and eases VRAM pressure.\n\nInference engine\nMany inference engines build in Flash Attention to run LLMs faster.", "relations": { "attention": { "label": "speeds up …", "note": "Flash Attention makes the Attention step faster and lighter." }, "transformer": { "label": "runs inside …", "note": "Transformers often use it to speed up their core work." }, "vram": { "label": "uses less …", "note": "It cuts the temporary memory needed during Attention." }, "llm-inference-engine": { "label": "is built into …", "note": "Inference engines use it to run LLMs faster." } } }, "zh": { "fullName": "快速注意力计算", "factExplain": "一种更省显存、计算更快的注意力方法。", "humanExplain": "它像外卖站分拣：不把包裹来回搬仓库，边扫边装车，少折腾就更快。\n\n它用于长上下文训练和推理，省显存，也常见于大模型加速。", "humanExplainDisplay": "它像==外卖站分拣==：\n不把包裹来回搬仓库，\n==边扫边装车==，\n少折腾就更快。\n\n它用于长上下文训练和推理，\n省显存，\n也常见于大模型加速。", "relationsNarrative": "Attention\n注意力是这步计算，Flash 把它做得更快、更省显存。\n\nTransformer\nTransformer 核心就是注意力，所以常靠它提速。\n\nVRAM\nFlash Attention 会减少中间过程的显存占用，缓解 VRAM 压力。\n\nLlm-inference-engine\n很多推理引擎会集成 Flash Attention，用来提升大模型运行表现。", "relations": { "attention": { "label": "加速…计算", "note": "它针对注意力这一步做了高效优化。" }, "transformer": { "label": "常用于…内部", "note": "Transformer 的核心计算常靠它提速。" }, "vram": { "label": "节省…占用", "note": "它能减少注意力计算时的显存压力。" }, "llm-inference-engine": { "label": "被…集成", "note": "推理引擎常集成它来提升运行效率。" } } } } }, { "id": "focal-loss", "name": "Focal Loss", "layer": "L2", "era": "2017", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "cross-entropy-loss" }, { "to": "object-detection" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Focal Loss", "factExplain": "A loss function for making the model focus on hard training examples.", "humanExplain": "Focal Loss is like a batting coach at practice. It stops cheering the kid hitting beach balls and helps the kid missing every pitch.\n\nIt makes easy training examples count less. You meet it in object detection. Background boxes can drown out real objects.", "humanExplainDisplay": "Focal Loss is like a ==batting coach== at practice.\nIt stops cheering the kid hitting beach balls\nand helps the kid ==missing every pitch==.\n\nIt makes easy training examples count less.\nYou meet it in object detection.\nBackground boxes can drown out real objects.", "relationsNarrative": "Cross-Entropy Loss\nFocal Loss reweights Cross-Entropy Loss to care less about easy examples.\n\nObject Detection\nFocal Loss helps Object Detection handle too many background examples.\n\nClassification\nFocal Loss still improves training for Classification.", "relations": { "cross-entropy-loss": { "label": "reworks …", "note": "It lowers the weight of easy examples in Cross-Entropy Loss." }, "object-detection": { "label": "helps …", "note": "Dense detection has far more background boxes than object boxes." }, "classification": { "label": "optimizes …", "note": "It still trains the model to choose the right class." } } }, "zh": { "fullName": "焦点损失", "factExplain": "一种让模型更关注难分类样本的损失函数。", "humanExplain": "焦点损失像急诊分诊：感冒号先放一边，专盯快恶化的病人。\n\n常用于目标检测，缓解正负样本失衡。", "humanExplainDisplay": "焦点损失像急诊分诊：\n==感冒号先放一边==，\n专盯==快恶化==的病人。\n\n常用于目标检测，\n缓解正负样本失衡。", "relationsNarrative": "Cross-Entropy Loss\n焦点损失是在交叉熵基础上加权改造。\n\nObject Detection\n它常用于目标检测中的样本不平衡问题。\n\nClassification\n它仍服务于分类任务的训练优化。", "relations": { "cross-entropy-loss": { "label": "改造…", "note": "它在交叉熵上降低简单样本权重。" }, "object-detection": { "label": "服务…", "note": "密集检测里背景样本常远多于目标。" }, "classification": { "label": "优化…", "note": "本质仍是在训练分类判断。" } } } } }, { "id": "foundation-model", "name": "Foundation-model", "layer": "L3", "era": "2021", "publishedAt": "2026-05-23T10:25:00Z", "relations": [ { "to": "pretraining" }, { "to": "fine-tuning" }, { "to": "llm" }, { "to": "multimodal" } ], "i18n": { "en": { "fullName": "Foundation Model", "factExplain": "A general AI model pretrained on huge data and adapted for many tasks.", "humanExplain": "A Foundation Model is like a giant box of LEGO. You can build a spaceship or a dog, but someone still needs a plan.\n\nIt sits under many AI apps. Teams tune it, add tools, or make it handle images and sound.", "humanExplainDisplay": "A Foundation Model is like a ==giant box of LEGO==.\nYou can build a spaceship or a dog,\nbut someone still needs ==a plan==.\n\nIt sits under many AI apps.\nTeams tune it, add tools,\nor make it handle images and sound.", "relationsNarrative": "Pretraining\nA Foundation Model gains broad skills from Pretraining.\n\nFine-tuning\nFine-tuning adapts a Foundation Model for a specific job.\n\nLLM\nAn LLM is a common Foundation Model for language work.\n\nMultimodal AI\nMultimodal expands a Foundation Model beyond text.", "relations": { "pretraining": { "label": "gets skills from …", "note": "Pretraining gives a Foundation Model broad skills before real use." }, "fine-tuning": { "label": "fits tasks through …", "note": "Fine-tuning adapts a Foundation Model for a specific job." }, "llm": { "label": "often appears as …", "note": "An LLM is a common Foundation Model for language work." }, "multimodal": { "label": "expands into …", "note": "Multimodal lets a Foundation Model handle more than text." } } }, "zh": { "fullName": "基础模型", "factExplain": "经过大规模预训练、可适配多类任务的通用模型。", "humanExplain": "基础模型像一块万能面团，能做面包、饺子、披萨，但还得看后厨怎么加工。\n\n它是很多 AI 应用的底座，之后可以微调、接工具或扩展成多模态系统。", "humanExplainDisplay": "基础模型像一块\n==万能面团==。\n能做面包、饺子、披萨，\n但还得看后厨怎么加工。\n\n它是很多 AI 应用的底座。\n之后可以微调、接工具，\n也可以扩成多模态系统。", "relationsNarrative": "Pretraining\nFoundation-model 通过 Pretraining 获得跨任务通用能力。\n\nFine-tuning\nFine-tuning 让 Foundation-model 进一步适配具体场景。\n\nLLM\nLLM 是 Foundation-model 在语言任务中的典型形态。\n\nMultimodal AI\nMultimodal 扩展了 Foundation-model 的输入和输出范围。", "relations": { "pretraining": { "label": "由…获得能力" }, "fine-tuning": { "label": "经…适配任务" }, "llm": { "label": "典型类型是…" }, "multimodal": { "label": "正扩展到…" } } } } }, { "id": "frame-problem", "name": "Frame Problem", "layer": "L1", "era": "1969", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "agent" }, { "to": "world-model" }, { "to": "nonmonotonic-reasoning" } ], "track": "history", "i18n": { "en": { "fullName": "Frame Problem", "factExplain": "The problem of knowing what changed after an action, and what stayed the same.", "humanExplain": "You move one cereal box in the pantry. Nobody checks if the fridge turned into a dragon.\n\nIt trips up common-sense reasoning. It also slows robots and Agents in messy places.", "humanExplainDisplay": "You move ==one cereal box==\nin the pantry.\nNobody checks\nif the ==fridge turned into a dragon==.\n\nIt trips up common-sense reasoning.\nIt also slows robots and Agents\nin messy places.", "relationsNarrative": "KR\nThe Frame Problem pushes KR to update only changed facts.\n\nAgent\nAn Agent gets slow if it recalculates everything before each move.\n\nWorld model\nA world model must know what changed and what stayed the same.\n\nNMR\nThe Frame Problem and NMR both deal with facts that can change.", "relations": { "knowledge-representation": { "label": "tests … choices", "note": "How you store facts decides what the AI can skip." }, "agent": { "label": "slows … planning", "note": "An Agent must know what stayed true after each move." }, "world-model": { "label": "tests … updates", "note": "A world model must update changes and leave steady facts alone." }, "nonmonotonic-reasoning": { "label": "is discussed with …", "note": "Both deal with reasoning when facts can change." } } }, "zh": { "fullName": "框架问题", "factExplain": "如何只更新相关事实、忽略无关变化的推理难题。", "humanExplain": "像武侠里你只挪了半步，旁观者没必要重算山河有没有移位；可 AI 常会把全世界都复盘一遍。\n\n它影响常识推理和行动规划，尤其卡机器人、智能体在复杂环境里的决策。", "humanExplainDisplay": "像武侠里你只挪了\n==半步==，\n旁观者没必要重算\n山河有没有移位；\n可 AI 常会把==全世界都复盘==一遍。\n\n它影响常识推理和行动规划，\n尤其卡机器人、\n智能体在复杂环境里的决策。", "relationsNarrative": "Knowledge Representation\n它是知识表示里的经典难题：怎样只更新该变的事实。\n\nAgent\n智能体做规划时，若样样重算，行动会又慢又笨。\n\nWorld Model\n世界模型需要知道什么变了、什么其实没变。\n\nNonmonotonic Reasoning\n两者都关心前提变化后，推理该如何调整。", "relations": { "knowledge-representation": { "label": "拷问…表达方式", "note": "怎么表示世界，决定能否少算无关事。" }, "agent": { "label": "限制…规划效率", "note": "智能体行动前，得判断哪些事实没变。" }, "world-model": { "label": "考验…更新能力", "note": "世界模型要改对变化，也别误伤常量。" }, "nonmonotonic-reasoning": { "label": "常与…一起讨论", "note": "两者都处理现实里会变的推理前提。" } } } } }, { "id": "frame-representation", "name": "Frame Representation", "layer": "L2", "era": "1974", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "symbolic-ai" }, { "to": "ontology" }, { "to": "expert-system" } ], "track": "history", "i18n": { "en": { "fullName": "Frame Representation", "factExplain": "A structure that stores knowledge with slots and default values.", "humanExplain": "Frame Representation is like a pizza order slip for facts. It has boxes for size and crust. Blank boxes get the house default.\n\nIt stores facts in slots. Expert systems and knowledge bases use it to look things up and fill common sense.", "humanExplainDisplay": "Frame Representation is like a ==pizza order slip== for facts.\nIt has boxes for size and crust.\nBlank boxes get the ==house default==.\n\nIt stores facts in slots.\nExpert systems and knowledge bases use it\nto look things up and fill common sense.", "relationsNarrative": "KR\nFrame Representation is a classic KR method with slots.\n\nSymbolic AI\nFrame Representation writes common sense as symbols for reasoning.\n\nOntology\nBoth organize concepts and relations. Frames feel more like slot templates.\n\nExpert System\nMany early Expert Systems used frames to store domain knowledge.", "relations": { "knowledge-representation": { "label": "is a … method", "note": "It puts common sense into named slots." }, "symbolic-ai": { "label": "supports … reasoning", "note": "It turns the world into symbols AI can reason with." }, "ontology": { "label": "models like …", "note": "Both sort concepts and links. Frames add slots." }, "expert-system": { "label": "stores knowledge for …", "note": "Early Expert Systems often used frames for domain knowledge." } } }, "zh": { "fullName": "框架表示", "factExplain": "用槽位和默认值表示对象或情境的知识结构。", "humanExplain": "框架表示像奶茶点单小票：甜度冰量有格子，没选就按默认做。\n\n用于早期专家系统，靠默认值把常识自动补齐。", "humanExplainDisplay": "框架表示像奶茶==点单小票==：\n甜度冰量有格子，\n没选就按==默认做==。\n\n用于早期专家系统，\n靠默认值把常识自动补齐。", "relationsNarrative": "Knowledge Representation\n它是用槽位组织知识的经典表示法。\n\nSymbolic AI\n它把常识写成可推理的符号结构。\n\nOntology\n两者都整理概念关系，框架更像带槽位的模板。\n\nExpert System\n许多早期专家系统用它存领域知识。", "relations": { "knowledge-representation": { "label": "作为…方法", "note": "用槽位把常识装进结构。" }, "symbolic-ai": { "label": "支撑…推理", "note": "把世界写成可推理符号。" }, "ontology": { "label": "近似…建模", "note": "都在梳理概念和关系。" }, "expert-system": { "label": "服务…存知识", "note": "早期系统常用框架存知识。" } } } } }, { "id": "framework", "name": "Framework", "layer": "L5", "sublayer": "architecture", "era": "2022", "publishedAt": "2026-05-23T10:40:00Z", "relations": [ { "to": "agent" }, { "to": "mcp" }, { "to": "rag" }, { "to": "api" } ], "i18n": { "en": { "fullName": "Application Framework", "factExplain": "A toolkit for organizing the parts and steps of an AI app.", "humanExplain": "A framework is like a meal kit for an AI app. It is not dinner, but it stops panic chopping.\n\nIt keeps the model and tools in order. It handles search and steps, so Agent or RAG apps are easier to build.", "humanExplainDisplay": "A framework is like a ==meal kit==\nfor an AI app.\nIt is not dinner,\nbut it stops ==panic chopping==.\n\nIt keeps the model and tools in order.\nIt handles search and steps,\nso Agent or RAG apps are easier to build.", "relationsNarrative": "Agent\nA framework runs the Agent's flow and tracks its state.\n\nMCP\nA framework can use MCP to connect to standard tools.\n\nRAG\nA framework can plug RAG into Q&A, search, and task flows.\n\nAPI\nA framework manages API calls in one place, so the app stays simpler.", "relations": { "agent": { "label": "organizes …", "note": "A framework runs the Agent's steps and tracks its state." }, "mcp": { "label": "connects through …", "note": "MCP lets a framework plug into standard tools." }, "rag": { "label": "wires in …", "note": "A framework can plug RAG into Q&A, search, and tasks." }, "api": { "label": "manages …", "note": "A framework manages API calls in one place." } } }, "zh": { "fullName": "应用框架", "factExplain": "用于组织 AI 应用开发流程和组件的工具体系。", "humanExplain": "框架像装修队先搭脚手架：别急着搬沙发，先让房子站得住。\n\n它常用于搭应用、接模型和管流程，让开发少从零开干。", "humanExplainDisplay": "框架像==装修队先搭脚手架==：\n别急着搬沙发，\n先让==房子站得住==。\n\n它常用于搭应用、\n接模型和管流程，\n让开发少从零开干。", "relationsNarrative": "Agent\nFramework 为 Agent 提供流程编排和状态管理。\n\nMCP\nFramework 可通过 MCP 接入标准化工具能力。\n\nRAG\nFramework 能把 RAG 接入问答、检索和任务流程。\n\nAPI\nFramework 统一管理 API 调用，降低应用复杂度。", "relations": { "agent": { "label": "组织…" }, "mcp": { "label": "连接…" }, "api": { "label": "整合…" } } } } }, { "id": "frontier-model-access-control", "name": "Frontier Model Access Control", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "frontier-model" }, { "to": "ai-regulation" }, { "to": "user-identity-verification-for-ai" }, { "to": "ai-biosecurity" } ], "track": "today", "i18n": { "en": { "fullName": "Frontier Model Access Control", "factExplain": "Rules that limit who can use frontier models and what they can do.", "humanExplain": "It is the velvet rope at the supercar counter. You can admire the car. To drive the 1,000-horsepower monster, show ID first.\n\nYou meet it in APIs, trials, and risky features. It helps stop abuse and the spread of dangerous know-how.", "humanExplainDisplay": "It is the ==velvet rope==\nat the supercar counter.\nYou can admire the car.\nTo drive the ==1,000-horsepower monster==,\nshow ID first.\n\nYou meet it in APIs, trials,\nand risky features.\nIt helps stop abuse\nand the spread of dangerous know-how.", "relationsNarrative": "Frontier model\nAccess control focuses on the strongest models with the highest risk.\n\nAI-regulation\nAI-regulation often asks platforms to review frontier model access by level.\n\nID Verification\nID Verification is the front door for access control.\n\nAI biosecurity\nLimits on biology-related abilities can lower misuse risk.", "relations": { "frontier-model": { "label": "limits access to …", "note": "The strongest models get the strictest gates." }, "ai-regulation": { "label": "follows … rules", "note": "Rules often require access levels and records." }, "user-identity-verification-for-ai": { "label": "checks users with …", "note": "First prove who you are, then get the right access." }, "ai-biosecurity": { "label": "blocks misuse in …", "note": "Biology-related tools may need extra locks." } } }, "zh": { "fullName": "前沿模型访问控制", "factExplain": "限制前沿模型使用者和能力范围的治理措施。", "humanExplain": "它像武馆门口的掌门把关：能看招式，想学杀招，先亮身份再过审。\n\n用于接口、试用和高危能力开放，防滥用、防扩散。", "humanExplainDisplay": "它像武馆门口的\n==掌门把关==：\n能看招式，\n想学==杀招==，\n先亮身份再过审。\n\n用于接口、试用，\n和高危能力开放，\n防滥用、防扩散。", "relationsNarrative": "Frontier Model\n访问控制专门盯住能力最强、风险最高的模型。\n\nAI Regulation\n监管常要求平台分级审核前沿模型访问。\n\nID Verification\n身份验证是落实访问控制的入口。\n\nAI Biosecurity\n限制生物相关能力，可降低滥用风险。", "relations": { "frontier-model": { "label": "限制…访问", "note": "只对高能力模型重点设门槛。" }, "ai-regulation": { "label": "响应…要求", "note": "法规常要求分级开放和审计。" }, "user-identity-verification-for-ai": { "label": "用…验人", "note": "先确认身份，才谈能用多少。" }, "ai-biosecurity": { "label": "防范…滥用", "note": "敏感生物能力常需额外限制。" } } } } }, { "id": "frontier-model", "name": "Frontier model", "layer": "L3", "era": "2023", "publishedAt": "2026-05-29T16:08:01.212Z", "relations": [ { "to": "foundation-model" }, { "to": "agi" }, { "to": "ai-regulation" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "Frontier model", "factExplain": "A large AI model with some of the strongest abilities today.", "humanExplain": "A frontier model is the kid at the science fair with the giant volcano. Everyone watches, and the teacher stands by the fire alarm.\n\nPeople use it to test what AI can do. It also starts safety and rule debates.", "humanExplainDisplay": "A frontier model is the kid at the ==science fair==\nwith the ==giant volcano==.\nEveryone watches,\nand the teacher stands by the fire alarm.\n\nPeople use it to test\nwhat AI can do.\nIt also starts safety\nand rule debates.", "relationsNarrative": "Foundation-model\nA frontier model is often one of the strongest foundation models.\n\nAGI\nPeople use frontier models to watch for signs of AGI.\n\nAI-regulation\nFrontier models often become a focus for AI-regulation.\n\nCompute-race\nA stronger frontier model can push a hotter compute-race.", "relations": { "foundation-model": { "label": "is often a top …", "note": "Frontier models are often the strongest foundation models." }, "agi": { "label": "is compared with …", "note": "People watch them for signs of more general intelligence." }, "ai-regulation": { "label": "triggers …", "note": "More power brings more attention from rule makers." }, "compute-race": { "label": "heats up …", "note": "The chase for the strongest model pushes demand for more computer power." } } }, "zh": { "fullName": "前沿模型", "factExplain": "能力处在当下最前列的大模型。", "humanExplain": "前沿模型像全班最会抢答的学霸，题还没念完，老师先担心它太会了。\n\n它常用于顶级产品和研究，也最容易引发安全、监管和算力竞争。", "humanExplainDisplay": "前沿模型像==全班最会抢答的学霸==，\n==题还没念完==，\n老师先担心它太会了。\n\n它常用于顶级产品和研究，\n也最容易引发安全、监管\n和算力竞争。", "relationsNarrative": "Foundation-model\nFrontier model 通常也是基础模型里能力最强、最受关注的一批。\n\nAGI\nFrontier model 常被拿来观察，是否正在逼近更通用的智能。\n\nAI-regulation\n因为能力和潜在风险都更高，Frontier model 更容易成为监管重点。\n\nCompute-race\n谁先做出更强的 Frontier model，往往会推动更激烈的算力竞争。", "relations": { "foundation-model": { "label": "通常属于…", "note": "前沿模型往往也是最强的一批基础模型。" }, "agi": { "label": "常被拿来对标…", "note": "很多人会用它观察是否逼近通用智能。" }, "ai-regulation": { "label": "最容易触发…", "note": "能力越靠前，越常进入监管视野。" }, "compute-race": { "label": "会加剧…", "note": "争夺最强模型常推动算力军备竞赛。" } } } } }, { "id": "fully-convolutional-network", "name": "FCN", "layer": "L3", "era": "2015", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "image-segmentation" }, { "to": "u-net" } ], "track": "history", "i18n": { "en": { "fullName": "Fully Convolutional Network", "factExplain": "A neural network that uses only convolution layers to predict every pixel.", "humanExplain": "FCN treats a photo like a coloring book page. It colors every tiny square: cat here, couch there.\n\nYou meet it in medical scans and self-driving car vision. It gives pixel positions, not one big label.", "humanExplainDisplay": "FCN treats a photo like a ==coloring book page==.\nIt colors ==every tiny square==:\ncat here, couch there.\n\nYou meet it in medical scans\nand self-driving car vision.\nIt gives pixel positions,\nnot one big label.", "relationsNarrative": "CNN\nFCN is a dense prediction version of CNN. It outputs a pixel map.\n\nImage Segmentation\nFCN helped image segmentation move from whole-image labels to pixel labels.\n\nU-Net\nU-Net follows the fully convolutional idea and improves detail blending.", "relations": { "cnn": { "label": "turns … into pixel prediction", "note": "FCN is a dense prediction version of CNN." }, "image-segmentation": { "label": "used for …", "note": "FCN became famous in semantic segmentation." }, "u-net": { "label": "inspired …", "note": "U-Net follows the fully convolutional path." } } }, "zh": { "fullName": "全卷积网络", "factExplain": "一种只用卷积层输出像素级预测的神经网络。", "humanExplain": "FCN 像给照片铺瓷砖：不只说有猫，还给每块砖标清猫和背景。\n\n用于医学影像和自动驾驶分割，输出像素级位置。", "humanExplainDisplay": "FCN 像给照片==铺瓷砖==：\n不只说有猫，\n还给==每块砖标清==\n猫和背景。\n\n用于医学影像\n和自动驾驶分割，\n输出像素级位置。", "relationsNarrative": "CNN\nFCN 是 CNN 的密集预测版本，能输出像素图。\n\nImage Segmentation\nFCN 让图像分割从整图判断走向逐像素标注。\n\nU-Net\nU-Net 继承全卷积思路，并强化细节融合。", "relations": { "cnn": { "label": "把…改成像素预测", "note": "FCN 是 CNN 的密集预测版。" }, "image-segmentation": { "label": "用于…", "note": "FCN 在语义分割中成名。" }, "u-net": { "label": "启发…", "note": "U-Net 延续了全卷积路线。" } } } } }, { "id": "function-call", "name": "Function-calling", "layer": "L4", "era": "2023", "publishedAt": "2026-05-23T09:45:00Z", "relations": [ { "to": "agent" }, { "to": "api" }, { "to": "mcp" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Function-calling", "factExplain": "A way for AI to call outside tools using neat, structured instructions.", "humanExplain": "Function-calling is like giving a chatbot a labeled remote. It can press the weather button, not just yell, “You got this!”\n\nIt sends neat instructions to other apps. Then AI can check weather or add a calendar event.", "humanExplainDisplay": "Function-calling is like giving a chatbot a ==labeled remote==.\nIt can ==press the weather button==,\nnot just yell, “You got this!”\n\nIt sends neat instructions to other apps.\nThen AI can check weather\nor add a calendar event.", "relationsNarrative": "Agent\nFunction-calling lets an Agent do the outside action it chose.\n\nAPI\nFunction-calling often triggers an API with structured details.\n\nMCP\nMCP makes the tool links behind Function-calling more standard.\n\nLLM\nFunction-calling helps an LLM move from writing text to doing tasks.", "relations": { "agent": { "label": "lets … take action", "note": "Function-calling lets an Agent do outside actions." }, "api": { "label": "runs through …", "note": "Function-calling often uses an API with clear inputs." }, "mcp": { "label": "can pair with …", "note": "MCP makes tool connections more standard for Function-calling." }, "llm": { "label": "extends … beyond text", "note": "Function-calling helps an LLM move from writing to doing." } } }, "zh": { "fullName": "函数调用", "factExplain": "让模型按结构化方式调用外部工具或程序的机制。", "humanExplain": "函数调用像给 AI 一只手：别光在群里出主意，真能去点外卖。\n\n它常用于查天气、下订单、改日程，是智能体动起来的接口。", "humanExplainDisplay": "函数调用像==给 AI 一只手==：\n别光在群里出主意，\n真能去==点外卖==。\n\n它常用于查天气、下订单、改日程，\n是智能体动起来的接口。", "relationsNarrative": "Agent\nAgent 借助 Function-call 执行模型决定的外部动作。\n\nAPI\nFunction-call 通常通过结构化参数触发 API 调用。\n\nMCP\nMCP 让 Function-call 背后的工具连接更加标准化。\n\nLLM\nFunction-call 扩展了 LLM 从生成文本到执行任务的能力。", "relations": { "agent": { "label": "让…能动手" }, "api": { "label": "通过…执行" }, "mcp": { "label": "可结合…" } } } } }, { "id": "game-ai", "name": "Game AI", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "ai-npc" }, { "to": "a-search" }, { "to": "minimax-search" }, { "to": "reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Game Artificial Intelligence", "factExplain": "AI tech that helps game characters, enemies, or systems make choices.", "humanExplain": "Game AI is the stage manager in a haunted house. It tells the zombie when to jump out, and when to miss on purpose.\n\nIt runs enemy behavior and pathfinding. It also helps pace a level, so the game feels alive.", "humanExplainDisplay": "Game AI is the ==stage manager== in a haunted house.\nIt tells the zombie when to jump out,\nand when to ==miss on purpose==.\n\nIt runs enemy behavior and pathfinding.\nIt also helps pace a level,\nso the game feels alive.", "relationsNarrative": "AI NPC\nGame AI often drives NPCs, so they seem to think.\n\nA* Search\nA* Search helps characters find a path around walls.\n\nMinimax Search\nMinimax Search helps game AI look ahead in board games.\n\nRL\nRL trains better play through lots of trial and error.", "relations": { "ai-npc": { "label": "drives … behavior", "note": "Game AI often shows up as NPC behavior." }, "a-search": { "label": "uses … for pathfinding", "note": "A* is a classic way to find a path on a map." }, "minimax-search": { "label": "uses … for game moves", "note": "Minimax helps AI plan moves in board games." }, "reinforcement-learning": { "label": "trains tactics with …", "note": "RL lets characters improve through trial and error." } } }, "zh": { "fullName": "游戏人工智能", "factExplain": "让游戏角色、敌人或系统做决策的 AI 技术。", "humanExplain": "游戏 AI 像剧本杀 DM：该吓就吓、该放水放水，还得装得真聪明。\n\n用于敌人行为、寻路和关卡调度，让游戏更鲜活。", "humanExplainDisplay": "游戏 AI 像==剧本杀 DM==：\n该吓就吓、该放水放水，\n还得装得真聪明。\n\n用于敌人行为、寻路，\n和关卡调度，\n让游戏更鲜活。", "relationsNarrative": "AI NPC\nAI NPC 是游戏 AI 最常见的落点，让角色像会思考。\n\nA* Search\nA* Search 常用于地图寻路，让角色会绕路追人。\n\nMinimax Search\nMinimax Search 适合棋类对弈，帮 AI 预判下一手。\n\nRL\nRL 能通过反复试错，训练更会玩的策略。", "relations": { "ai-npc": { "label": "驱动…行为", "note": "游戏 AI 常落在 NPC 行为上。" }, "a-search": { "label": "用…做寻路", "note": "A* 是经典地图寻路办法。" }, "minimax-search": { "label": "用…做对弈决策", "note": "Minimax 常用于棋类博弈。" }, "reinforcement-learning": { "label": "用…训练策略", "note": "RL 可让角色从试错中变强。" } } } } }, { "id": "gan", "name": "GAN", "layer": "L3", "era": "2014", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "deepfake" }, { "to": "diffusion" }, { "to": "adversarial-example" } ], "track": "history", "i18n": { "en": { "fullName": "Generative Adversarial Network", "factExplain": "A generative model where two neural networks learn by competing.", "humanExplain": "A GAN is like a fake-ID kid and a sharp-eyed bouncer. The kid gets better because the bouncer keeps saying, “Nice try.”\n\nYou meet it in fake faces and image editing. It also helped deepfakes look real.", "humanExplainDisplay": "A GAN is like a ==fake-ID kid==\nand a ==sharp-eyed bouncer==.\nThe kid gets better\nbecause the bouncer keeps saying,\n“Nice try.”\n\nYou meet it in fake faces\nand image editing.\nIt also helped deepfakes\nlook real.", "relationsNarrative": "Generative Model\nGAN is a generative model. It learns to make realistic samples.\n\nDeepfake\nGAN was often used for face swaps and fake human faces.\n\nDiffusion\nGAN and Diffusion can both make images, but they train differently.\n\nAdv Example\nBoth use adversarial ideas, but they aim at different goals.", "relations": { "generative-model": { "label": "is a kind of …", "note": "GAN is a classic type of generative model." }, "deepfake": { "label": "often powers …", "note": "Realistic fake faces helped deepfakes grow." }, "diffusion": { "label": "competes with …", "note": "Both can make images, but they learn in different ways." }, "adversarial-example": { "label": "shares adversarial ideas with …", "note": "Both use a contest idea, but for different goals." } } }, "zh": { "fullName": "生成对抗网络", "factExplain": "一种让两个网络对抗学习的生成模型。", "humanExplain": "GAN 像夜市套圈摊上俩老板斗法：一个专门做高仿，一个专门挑毛病，越较劲越能以假乱真。\n\n常用于生成人脸、修图和图像合成，也推动了深伪内容变逼真。", "humanExplainDisplay": "GAN 像夜市套圈摊上\n俩老板==斗法==：\n一个专门做高仿，\n一个专门==挑毛病==，\n越较劲越能以假乱真。\n\n常用于生成人脸、修图\n和图像合成，\n也推动了深伪内容\n变逼真。", "relationsNarrative": "Generative Model\nGAN 是生成模型的一种，用来学习造出逼真样本。\n\nDeepfake\nGAN 早期常被用于换脸和拟真人像生成。\n\nDiffusion\n它和 Diffusion 都能生图，但训练方式不同。\n\nAdversarial-example\n两者都带有对抗思路，但目标并不一样。", "relations": { "generative-model": { "label": "属于…一类", "note": "它是经典生成模型代表之一。" }, "deepfake": { "label": "常被…采用", "note": "高拟真人脸生成推动深伪发展。" }, "diffusion": { "label": "与…同台竞争", "note": "二者都能生成图像，但路线不同。" }, "adversarial-example": { "label": "同属对抗思路", "note": "都与对抗训练和脆弱性相关。" } } } } }, { "id": "gaussian-mixture-model", "name": "GMM", "layer": "L3", "era": "1960s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "expectation-maximization" }, { "to": "unsupervised-learning" }, { "to": "hidden-markov-model" }, { "to": "generative-model" } ], "track": "history", "i18n": { "en": { "fullName": "Gaussian Mixture Model", "factExplain": "A probability model that describes data as several bell-shaped groups mixed together.", "humanExplain": "GMM is like sorting laundry after a sleepover. One basket looks messy, but each kid has a usual size.\n\nIt helps find hidden groups in mixed data. It also estimates density and spots odd cases.", "humanExplainDisplay": "GMM is like sorting ==laundry after a sleepover==.\nOne basket looks messy,\nbut each kid has a ==usual size==.\n\nIt helps find hidden groups in mixed data.\nIt also estimates density and spots odd cases.", "relationsNarrative": "EM\nGMM often uses EM to guess group chances and update the curve settings.\n\nUnsupervised Learning\nGMM is a classic unsupervised method for finding hidden clusters.\n\nHMM\nAn HMM can use GMM to model continuous outputs from each state.\n\nGenerative Model\nGMM learns how data is made, then uses that model for clustering or decisions.", "relations": { "expectation-maximization": { "label": "is often trained by …", "note": "EM repeats two steps: guess the groups, then update the curves." }, "unsupervised-learning": { "label": "is a classic … method", "note": "It finds hidden groups in data with no labels." }, "hidden-markov-model": { "label": "can serve as outputs for …", "note": "An HMM can use it to describe the output from each state." }, "generative-model": { "label": "is a type of …", "note": "It learns how data is shaped, then uses that shape to decide or create." } } }, "zh": { "fullName": "高斯混合模型", "factExplain": "用多个高斯分布共同表示数据分布的概率模型。", "humanExplain": "它像老中医把脉分证：表面都像一个毛病，细看却是几路体质，各有自己的正常范围。\n\n常用于聚类、密度估计和异常检测，适合数据里混着多拨来源的情况。", "humanExplainDisplay": "它像老中医把脉分证：\n表面都像一个毛病，\n细看却是几路==体质==，\n各有自己的==正常范围==。\n\n常用于聚类、密度估计\n和异常检测，\n适合数据里混着多拨来源的情况。", "relationsNarrative": "Expectation-maximization\n它常用 EM 交替估计分组概率和分布参数。\n\nUnsupervised-learning\n它是经典无监督方法，常拿来发现隐藏簇。\n\nHidden-markov-model\n在连续观测场景里，它常被用作 HMM 的发射分布。\n\nGenerative-model\n它先刻画数据怎么来，再据此做聚类或判别。", "relations": { "expectation-maximization": { "label": "常用…来训练", "note": "它常靠 EM 反复估参数和分组。" }, "unsupervised-learning": { "label": "属于…典型方法", "note": "常在没标签数据里找隐藏群体。" }, "hidden-markov-model": { "label": "可作为…的观测层", "note": "HMM 常用它描述每个状态的输出。" }, "generative-model": { "label": "属于…一类", "note": "它先学数据分布，再做判断或生成。" } } } } }, { "id": "gaussian-process", "name": "GP", "layer": "L3", "era": "1970s", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "kernel-method" }, { "to": "regression" }, { "to": "model-uncertainty" } ], "track": "history", "i18n": { "en": { "fullName": "Gaussian Process", "factExplain": "A model that uses Gaussian distributions to describe likely functions.", "humanExplain": "A Gaussian Process is like a weather app for a squiggly line. It gives a best guess, plus a “don’t bet your lunch money” zone.\n\nPeople use it for small-data regression and Bayesian optimization. It also shows how unsure each guess is.", "humanExplainDisplay": "A Gaussian Process is like a ==weather app for a squiggly line==.\nIt gives a best guess,\nplus a ==“don’t bet your lunch money” zone==.\n\nPeople use it for small-data regression\nand Bayesian optimization.\nIt also shows how unsure each guess is.", "relationsNarrative": "Kernel Method\nThe kernel defines input similarity in a Gaussian Process.\n\nRegression\nA Gaussian Process often does regression and gives a prediction range.\n\nModel uncertainty\nA Gaussian Process directly shows how uncertain its prediction is.", "relations": { "kernel-method": { "label": "measures similarity with …", "note": "The kernel decides which inputs can affect each other." }, "regression": { "label": "is often used for …", "note": "Gaussian Process regression gives a prediction range too." }, "model-uncertainty": { "label": "shows … directly", "note": "It naturally says how unsure each prediction is." } } }, "zh": { "fullName": "高斯过程", "factExplain": "一种以高斯分布刻画函数的非参数模型。", "humanExplain": "高斯过程像老中医把脉：不只说病因，还会补一句“有七成把握”。\n\n适合小数据回归和贝叶斯优化；能报不确定性。", "humanExplainDisplay": "高斯过程像\n==老中医把脉==：\n不只说病因，\n还会补一句==有七成把握==。\n\n适合小数据回归和贝叶斯优化；\n能报\n不确定性。", "relationsNarrative": "Kernel Method\n核函数定义输入相似度，是高斯过程的核心。\n\nRegression\n高斯过程常做回归，并给出预测区间。\n\nModel Uncertainty\n它会直接报告预测结果有多不确定。", "relations": { "kernel-method": { "label": "用…量相似", "note": "核函数决定哪些输入会互相影响。" }, "regression": { "label": "常用于…", "note": "高斯过程回归会同时给出预测区间。" }, "model-uncertainty": { "label": "直接表达…", "note": "它天生会报告预测有多没把握。" } } } } }, { "id": "gemini-omni-flash", "name": "Gemini Omni Flash", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "gemini" }, { "to": "multimodal" }, { "to": "streaming-multimodal-model" } ], "track": "today", "i18n": { "en": { "fullName": "Gemini Omni Flash", "factExplain": "A Google Gemini model built for fast, low-lag multimodal chats.", "humanExplain": "Gemini Omni Flash is like a drive-thru pro on headset duty. It hears you. It sees the screen. It answers before the fries get cold.\n\nIt is built for live voice helpers and photo questions. It helps apps handle sound and images with less waiting.", "humanExplainDisplay": "Gemini Omni Flash is like a ==drive-thru pro==\non headset duty.\nIt hears you.\nIt sees the screen.\nIt answers before the ==fries get cold==.\n\nIt is built for live voice helpers\nand photo questions.\nIt helps apps handle sound and images\nwith less waiting.", "relationsNarrative": "Gemini\nGemini Omni Flash is the fast multimodal model in the Gemini family.\n\nMultimodal AI\nIt handles text, images, and sound in one model.\n\nLive Multimodal\nIts low lag makes live voice and video chats smoother.", "relations": { "gemini": { "label": "belongs to …", "note": "It is a fast multimodal branch of the Gemini family." }, "multimodal": { "label": "boosts …", "note": "Its main trick is handling more than one input type at once." }, "streaming-multimodal-model": { "label": "aims at …", "note": "Low lag makes voice and video chats feel more natural." } } }, "zh": { "fullName": "Gemini 全模态快速版", "factExplain": "Google 面向低延迟多模态交互的 Gemini 模型。", "humanExplain": "Gemini Omni Flash像电竞陪玩开全麦：看屏、听麦、秒回话，反应快到不掉帧。\n\n适合实时语音助手、拍照问答，支撑低延迟多模态应用。", "humanExplainDisplay": "Gemini Omni Flash\n像==电竞陪玩开全麦==：\n看屏、听麦、秒回话，\n==反应快到不掉帧==。\n\n适合实时语音助手、\n拍照问答，\n支撑低延迟多模态应用。", "relationsNarrative": "Gemini\n它是 Gemini 系列里主打快速全模态交互的模型。\n\nMultimodal AI\n它把文字、图像、语音等输入放在同一套模型里处理。\n\nLive Multimodal\n低延迟能力让实时语音和视频对话更顺。", "relations": { "gemini": { "label": "属于…家族", "note": "它是 Gemini 系列的快速多模态分支。" }, "multimodal": { "label": "强化…能力", "note": "核心卖点是同时处理多种输入。" }, "streaming-multimodal-model": { "label": "面向…体验", "note": "低延迟让语音视频互动更自然。" } } } } }, { "id": "gemini", "name": "Gemini", "layer": "L3", "era": "2023", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "llm" }, { "to": "gemma" }, { "to": "chatgpt" } ], "track": "history", "i18n": { "en": { "fullName": "Google Gemini", "factExplain": "Google’s AI model family for words, images, audio, and more.", "humanExplain": "Gemini is Google’s group-project overachiever. Show it a photo or hand it text, and it says, “I’ll make the slides.”\n\nYou meet it in Google chat and search. It also sits behind work tools and coding tools from Google.", "humanExplainDisplay": "Gemini is Google’s ==group-project overachiever==.\nShow it a photo or hand it text,\nand it says, ==“I’ll make the slides.”==\n\nYou meet it in Google chat and search.\nIt also sits behind work tools\nand coding tools from Google.", "relationsNarrative": "Multimodal AI\nGemini is a major example of Multimodal AI in use.\n\nLLM\nGemini builds on LLM skills and adds more input types.\n\nGemma\nGemini and Gemma both come from Google, but they have different jobs.\n\nChatGPT\nPeople often compare Gemini with ChatGPT for skills, apps, and user experience.", "relations": { "multimodal": { "label": "is a major … model", "note": "Gemini is a real-world example of Multimodal AI." }, "llm": { "label": "extends …", "note": "It starts with language skills and adds other input types." }, "gemma": { "label": "shares Google roots with …", "note": "Both come from Google, but they serve different jobs." }, "chatgpt": { "label": "competes with …", "note": "People compare their skills, apps, and user experience." } } }, "zh": { "fullName": "谷歌多模态大模型系列", "factExplain": "谷歌推出的一系列多模态大模型。", "humanExplain": "Gemini 像班里那个全能课代表：老师一句话，他既能看题、听话、读图，还能顺手把活都接过去。\n\n常用于聊天、搜索、办公和开发，是谷歌多类 AI 产品的重要模型底座。", "humanExplainDisplay": "Gemini 像班里那个\n==全能课代表==：\n老师一句话，他既能看题、\n听话、读图，还能顺手\n把活都==接过去==。\n\n常用于聊天、搜索、\n办公和开发，是谷歌\n多类 AI 产品的重要模型底座。", "relationsNarrative": "Multimodal\n它是多模态能力的代表性模型之一。\n\nLLM\n它建立在语言模型能力上，并扩展到多模态输入输出。\n\nGemma\n两者都出自谷歌，但一个偏前沿产品，一个偏开放模型。\n\nChatGPT\n它常被拿来和 ChatGPT 比能力、体验与生态。", "relations": { "multimodal": { "label": "属于…代表模型", "note": "它是多模态能力的典型落地。" }, "llm": { "label": "是…的一种扩展", "note": "在语言模型上加入多模态理解。" }, "gemma": { "label": "与…同属谷歌系", "note": "两者都出自谷歌，但定位不同。" }, "chatgpt": { "label": "常与…对比", "note": "它们是主流通用助手的竞品。" } } } } }, { "id": "gemma", "name": "Gemma", "layer": "L3", "era": "2024", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "open-weights" }, { "to": "local-llm" }, { "to": "fine-tuning" }, { "to": "small-language-model" } ], "track": "today", "i18n": { "en": { "fullName": "Gemma", "factExplain": "A lightweight Google model family with open weights for commercial use.", "humanExplain": "Gemma is the mini toolbox in your kitchen drawer. It will not build a house, but it can save Saturday.\n\nPeople run it on their own machines for more control. Teams tune it for special jobs, or put it on smaller devices.", "humanExplainDisplay": "Gemma is the ==mini toolbox==\nin your kitchen drawer.\nIt will not build a house,\nbut it can ==save Saturday==.\n\nPeople run it on their own machines\nfor more control.\nTeams tune it for special jobs,\nor put it on smaller devices.", "relationsNarrative": "Open weights\nGemma usually comes with open weights, so people can download and run it.\n\nLocal-LLM\nGemma is small enough for many local and offline setups.\n\nFine-tuning\nTeams often fine-tune Gemma for their own jobs.\n\nSLM\nGemma is a well-known SLM family.", "relations": { "open-weights": { "label": "usually ships with …", "note": "Open weights make Gemma easy to download and run yourself." }, "local-llm": { "label": "often used for …", "note": "Its smaller size helps it run on local machines." }, "fine-tuning": { "label": "can be adapted with …", "note": "Teams fine-tune Gemma for their own jobs." }, "small-language-model": { "label": "is an example of …", "note": "Gemma is a well-known SLM family." } } }, "zh": { "fullName": "谷歌开源轻量模型系列", "factExplain": "Google 推出的可商用开放权重轻量模型家族。", "humanExplain": "像楼下便利店那把顺手的小折叠伞：不占地方，真下雨时一撑就能顶事。\n\n适合本地部署、继续微调和边缘设备使用，想自己掌控模型时常会选它。", "humanExplainDisplay": "像楼下便利店那把顺手的\n==小折叠伞==：\n不占地方，\n真下雨时一撑就能==顶事==。\n\n适合本地部署、\n继续微调和边缘设备使用，\n想自己掌控模型时常会选它。", "relationsNarrative": "Open-weights\n它通常以开放权重形式发布，方便社区下载和部署。\n\nLocal-LLM\nGemma 因体量较轻，常被拿来做本地运行和离线使用。\n\nFine-tuning\n很多团队会在它基础上继续微调，做行业或任务适配。\n\nSmall-language-model\n它是近年小语言模型路线里较有代表性的一个系列。", "relations": { "open-weights": { "label": "通常以…发布", "note": "它靠开放权重方便下载与自部署。" }, "local-llm": { "label": "常被用于…", "note": "参数更轻，适合本地跑起来。" }, "fine-tuning": { "label": "可继续做…", "note": "很多团队拿它做垂直场景适配。" }, "small-language-model": { "label": "属于…代表", "note": "它常被视作小模型路线代表之一。" } } } } }, { "id": "general-problem-solver", "name": "GPS", "layer": "L4", "era": "1959", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "physical-symbol-system-hypothesis" }, { "to": "automated-planning" }, { "to": "symbolic-ai" } ], "track": "history", "i18n": { "en": { "fullName": "General Problem Solver", "factExplain": "An early AI program that solved problems by closing gaps to a goal.", "humanExplain": "GPS is like copying a LEGO castle from the box. It spots a missing tower, adds bricks, then hunts the next gap.\n\nIn early AI, it used this gap-closing method to solve tasks. It helped inspire Planning, where machines work step by step.", "humanExplainDisplay": "GPS is like copying a ==LEGO castle== from the box.\nIt spots a ==missing tower==,\nadds bricks,\nthen hunts the next gap.\n\nIn early AI,\nit used this gap-closing method to solve tasks.\nIt helped inspire Planning,\nwhere machines work step by step.", "relationsNarrative": "Physical Symbol System Hypothesis\nGPS was often used as proof that symbol systems could show intelligence.\n\nPlanning\nGPS used means-end analysis, which shaped automated planning.\n\nSymbolic AI\nGPS was one of the early showcase programs of Symbolic AI.", "relations": { "physical-symbol-system-hypothesis": { "label": "supports …", "note": "GPS showed how symbol rules could look smart." }, "automated-planning": { "label": "inspired …", "note": "Means-end analysis shaped later planning research." }, "symbolic-ai": { "label": "belongs to …", "note": "GPS was an early Symbolic AI program." } } }, "zh": { "fullName": "General Problem Solver（通用问题求解器）", "factExplain": "一种用手段-目的分析求解问题的早期 AI 程序。", "humanExplain": "GPS 像外卖骑手找小区门：先看离目标差几步，再一关关补路线。\n\n它启发符号规划，让机器在任务求解中逐步补差距。", "humanExplainDisplay": "GPS 像==外卖骑手==找小区门：\n先看离目标差几步，\n再一关关\n==补路线==。\n\n它启发符号规划，\n让机器在任务求解中\n逐步补差距。", "relationsNarrative": "Physical Symbol System Hypothesis\n它常被当作符号系统能显智能的例证。\n\nPlanning\n它的手段-目的分析影响了自动规划。\n\nSymbolic AI\n它是早期符号主义的代表程序之一。", "relations": { "physical-symbol-system-hypothesis": { "label": "支撑…", "note": "它被用来说明符号操作能显智能。" }, "automated-planning": { "label": "启发…", "note": "手段-目的分析影响规划研究。" }, "symbolic-ai": { "label": "属于…", "note": "它是符号主义早期代表程序。" } } } } }, { "id": "generalized-advantage-estimation", "name": "GAE", "layer": "L2", "era": "2015", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "policy-gradient" }, { "to": "actor-critic" }, { "to": "temporal-difference-learning" }, { "to": "proximal-policy-optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Generalized Advantage Estimation", "factExplain": "A low-noise method for estimating how much an action helped a policy.", "humanExplain": "GAE is like grading a taco run, not one missed turn. Check the whole ride before you judge.\n\nIt estimates how much each action helped, with less noise. In Policy Gradient and robot training, it makes updates less jumpy.", "humanExplainDisplay": "GAE is like grading a ==taco run==,\nnot one missed turn.\nCheck the ==whole ride== before you judge.\n\nIt estimates how much each action helped,\nwith less noise.\nIn Policy Gradient and robot training,\nit makes updates less jumpy.", "relationsNarrative": "Policy Gradient\nGAE gives Policy Gradient a steadier advantage estimate.\n\nActor-Critic\nThe critic estimates value, and GAE makes the advantage signal steadier.\n\nTD Learning\nGAE uses multi-step TD errors to balance bias and variance.\n\nPPO\nPPO often uses GAE to calculate each action's advantage.", "relations": { "policy-gradient": { "label": "estimates advantage for …", "note": "Before a policy update, GAE checks whether an action paid off." }, "actor-critic": { "label": "trains with …", "note": "The critic gives values, and GAE makes the advantage steadier." }, "temporal-difference-learning": { "label": "smooths rewards with …", "note": "Multi-step TD errors let it balance bias and variance." }, "proximal-policy-optimization": { "label": "often used in …", "note": "PPO often uses GAE to score each step's action." } } }, "zh": { "fullName": "Generalized Advantage Estimation，广义优势估计", "factExplain": "一种估计策略优势的低方差方法。", "humanExplain": "GAE像篮球赛后看正负值：别只记一球打铁，要算整段在场赢没赢。\n\n用于策略梯度和机器人训练，降低方差，让更新更稳。", "humanExplainDisplay": "GAE像篮球赛后看正负值：\n别只记==一球打铁==，\n要算整段在场\n==赢没赢==。\n\n用于策略梯度和机器人训练，\n降低方差，\n让更新更稳。", "relationsNarrative": "Policy Gradient\n它为策略梯度提供更稳的优势估计。\n\nActor-Critic\n评论家估价值，它负责把优势信号算稳。\n\nTD Learning\n它用多步 TD 误差折中偏差和方差。\n\nPPO\nPPO 常用它来计算每步动作的优势。", "relations": { "policy-gradient": { "label": "为…估计优势", "note": "策略更新前，先判断动作值不值。" }, "actor-critic": { "label": "配合…训练", "note": "评论家给价值，它把优势算稳。" }, "temporal-difference-learning": { "label": "借…平滑奖励", "note": "用多步 TD 误差折中偏差和方差。" }, "proximal-policy-optimization": { "label": "常用于…训练", "note": "PPO 常靠它提供优势信号。" } } } } }, { "id": "generative-architecture-design", "name": "Generative Architecture Design", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "ai-accelerated-prototyping" }, { "to": "multimodal" } ], "track": "today", "i18n": { "en": { "fullName": "Generative Architecture Design", "factExplain": "Using generative AI to create many building design ideas automatically.", "humanExplain": "It is like a tireless architect dumping floor plans on your kitchen table. Move the bathroom, and boom, a fresh stack appears.\n\nArchitects use it early to explore ideas fast. People still choose the final design.", "humanExplainDisplay": "It is like a tireless architect\n dumping ==floor plans== on your kitchen table.\nMove the bathroom,\nand boom,\na ==fresh stack== appears.\n\nArchitects use it early\nto explore ideas fast.\nPeople still choose the final design.", "relationsNarrative": "Generative Model\nGenerative Architecture Design uses a Generative Model to make design options.\n\nAI Prototyping\nGenerative Architecture Design speeds up early testing of building ideas.\n\nMultimodal AI\nGenerative Architecture Design can use text, sketches, and images together.", "relations": { "generative-model": { "label": "is driven by …", "note": "A Generative Model usually creates the design options." }, "ai-accelerated-prototyping": { "label": "speeds up …", "note": "It makes early design testing much faster." }, "multimodal": { "label": "uses … input", "note": "It can work with text, sketches, and images." } } }, "zh": { "fullName": "生成式建筑设计", "factExplain": "用生成式 AI 自动提出建筑设计方案的方法。", "humanExplain": "像老裁缝赶制样衣，腰收一寸、袖改一截，转身又能铺出一摞新版型让你挑。\n\n常用于建筑前期方案探索，帮助快速试错，最终拍板仍靠人。", "humanExplainDisplay": "像老裁缝赶制==样衣==，\n腰收一寸、袖改一截，\n转身又能铺出\n一摞==新版型==让你挑。\n\n常用于建筑前期\n方案探索，\n帮助快速试错，\n最终拍板仍靠人。", "relationsNarrative": "Generative Model\n它通常由生成模型驱动，自动产出多个设计方案。\n\nAI Prototyping\n它把建筑方案的早期试错和打样过程加快了。\n\nMultimodal AI\n它常结合文字、草图和图像一起理解设计需求。", "relations": { "generative-model": { "label": "由…驱动生成", "note": "底层通常靠生成模型产出方案。" }, "ai-accelerated-prototyping": { "label": "加速…过程", "note": "能把早期方案试错明显提速。" }, "multimodal": { "label": "常结合…输入", "note": "可同时处理文本、草图和图像。" } } } } }, { "id": "generative-games", "name": "Generative Games", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "ai-npc" }, { "to": "world-model" }, { "to": "llm-game-benchmark" } ], "track": "today", "i18n": { "en": { "fullName": "Generative Games", "factExplain": "Games that use generative AI to create content and play moments on the fly.", "humanExplain": "Generative games are like a board game with a mischievous referee. Pick the safe path, and it adds a trapdoor.\n\nThey create fresh maps and story turns while you play. Characters can answer back, so the game feels less scripted.", "humanExplainDisplay": "Generative games are like a board game\nwith a ==mischievous referee==.\nPick the ==safe path==,\nand it adds a trapdoor.\n\nThey create fresh maps\nand story turns while you play.\nCharacters can answer back,\nso the game feels less scripted.", "relationsNarrative": "Generative Model\nGenerative games use a Generative Model to create content in real time.\n\nAI NPC\nGenerative games can make an AI NPC chat and react with more life.\n\nWorld model\nA World model helps generated game content stay consistent.\n\nLLM Game Benchmark\nAn LLM Game Benchmark can test how well a model plans and adapts.", "relations": { "generative-model": { "label": "built on …", "note": "The generative model decides how much the game can change." }, "ai-npc": { "label": "helps … improvise", "note": "Generative AI makes NPC talk and react more naturally." }, "world-model": { "label": "builds worlds with …", "note": "A world model helps new content stay consistent." }, "llm-game-benchmark": { "label": "can be tested by …", "note": "Game tasks can test how well an AI plans and adapts." } } }, "zh": { "fullName": "生成式游戏", "factExplain": "用生成式 AI 动态生成游戏内容与体验的游戏形态。", "humanExplain": "像剧本杀主持人会临场加戏：你刚想走老路，它就给你塞个新角色和新岔口。\n\n常用于生成地图、剧情和角色互动，让游戏更耐玩，也更难被写死。", "humanExplainDisplay": "像剧本杀主持人\n会临场==加戏==：\n你刚想走老路，\n它就给你塞个\n==新岔口==。\n\n常用于生成地图、剧情\n和角色互动，\n让游戏更耐玩，\n也更难被写死。", "relationsNarrative": "Generative Model\n它建立在生成模型之上，靠模型实时产出内容。\n\nAI NPC\n生成式游戏常让 NPC 更能聊、更会即兴反应。\n\nWorld Model\n世界模型能帮助生成内容前后连贯、不乱套。\n\nLLM Game Benchmark\n游戏环境也常被拿来测试模型的规划与应变。", "relations": { "generative-model": { "label": "建立在…之上", "note": "底层生成能力决定内容变化空间。" }, "ai-npc": { "label": "让…更会演", "note": "生成式能力能让 NPC 对话更活。" }, "world-model": { "label": "可结合…构世界", "note": "世界模型让生成结果更连贯。" }, "llm-game-benchmark": { "label": "可被…测试", "note": "游戏任务常被用来评估智能体能力。" } } } } }, { "id": "generative-model", "name": "Generative Model", "layer": "L2", "era": "1980", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "diffusion" }, { "to": "deepfake" }, { "to": "foundation-model" } ], "track": "history", "i18n": { "en": { "fullName": "Generative Model", "factExplain": "A model for learning data patterns and creating new content.", "humanExplain": "It is like a school cafeteria cook with every lunch combo memorized. Ask for “space tacos,” and you get a fresh plate.\n\nIt can make text, images, music, and video. It is the main engine behind AIGC.", "humanExplainDisplay": "It is like a ==school cafeteria cook==\nwith every lunch combo memorized.\nAsk for ==“space tacos,”==\nand you get a fresh plate.\n\nIt can make text, images, music, and video.\nIt is the main engine behind AIGC.", "relationsNarrative": "LLM\nLLMs are a main text branch of generative models.\n\nDiffusion\nDiffusion is a leading way to generate images and videos.\n\nDeepfake\nGenerative models make Deepfake content look real, so misuse gets easier.\n\nFoundation-model\nMany Foundation-models are generative models at heart.", "relations": { "llm": { "label": "includes …", "note": "An LLM is a key type of generative model." }, "diffusion": { "label": "has the … route", "note": "Diffusion is a leading way to generate images and videos." }, "deepfake": { "label": "powers … generation", "note": "Realistic generation can make fake content easier." }, "foundation-model": { "label": "forms many …", "note": "Many foundation models are generative models at heart." } } }, "zh": { "fullName": "生成模型（Generative Model）", "factExplain": "学习数据分布并生成新内容的一类模型。", "humanExplain": "它像网购店里的万能客服，见过太多款式后，你一句需求，它就能现拼出一版给你看。\n\n能生成文字、图片、音乐和视频，也是如今 AIGC 的核心能力。", "humanExplainDisplay": "它像网购店里的==万能客服==，\n见过太多款式后，\n你一句需求，\n它就能==现拼出一版==给你看。\n\n能生成文字、图片、\n音乐和视频，\n也是如今 AIGC 的核心能力。", "relationsNarrative": "LLM\n大语言模型是生成模型在文本领域的一条主线。\n\nDiffusion\n扩散模型是生成模型在图像与视频里的代表方案。\n\nDeepfake\n生成模型让伪造内容更逼真，也放大了滥用风险。\n\nFoundation-model\n很多基础模型本质上都属于生成模型。", "relations": { "llm": { "label": "包含…这一支", "note": "大语言模型是生成模型的重要类型。" }, "diffusion": { "label": "有…这条路线", "note": "扩散模型是生成图像视频的代表方案。" }, "deepfake": { "label": "支撑…生成", "note": "能生成逼真内容，也带来伪造风险。" }, "foundation-model": { "label": "构成…主力", "note": "很多基础模型本质上都是生成模型。" } } } } }, { "id": "genetic-algorithm", "name": "GA", "layer": "L2", "era": "1975", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "optimization" }, { "to": "heuristic-search" }, { "to": "hyperparameter-optimization" }, { "to": "no-free-lunch-theorem" } ], "track": "history", "i18n": { "en": { "fullName": "Genetic Algorithm", "factExplain": "GA keeps good tries, mixes them, and adds random changes.", "humanExplain": "GA is like a toy-car race in a garage. Keep the fast cars, swap their parts, and sometimes add a weird rocket.\n\nIt hunts a huge pile of choices for a better answer. It helps with routes and schedules. It also picks AI settings.", "humanExplainDisplay": "GA is like a ==toy-car race== in a garage.\nKeep the fast cars,\nswap their parts,\nand sometimes add a ==weird rocket==.\n\nIt hunts a huge pile of choices\nfor a better answer.\nIt helps with routes and schedules.\nIt also picks AI settings.", "relationsNarrative": "Optimization\nGA is an optimization method for finding better answers in a huge space.\n\nHeuristic Search\nGA uses evolution-style trial and error, not a full check of every answer.\n\nHPO\nGA can automatically try different hyperparameter mixes.\n\nNo Free Lunch Theorem\nThe No Free Lunch Theorem says GA is not best for every problem.", "relations": { "optimization": { "label": "used for …", "note": "It searches a huge space for a better plan." }, "heuristic-search": { "label": "is a kind of …", "note": "It tries smart guesses instead of checking every answer." }, "hyperparameter-optimization": { "label": "can be used for …", "note": "It can automatically test mixes of model settings." }, "no-free-lunch-theorem": { "label": "is limited by …", "note": "No search method is best for every problem." } } }, "zh": { "fullName": "遗传算法（Genetic Algorithm）", "factExplain": "用选择、交叉、变异迭代搜索最优解的算法。", "humanExplain": "遗传算法像方案相亲：优胜者配对生娃，偶尔基因突变冒黑马。\n\n常用于路径、排班、调参，在巨大解空间里找较优解。", "humanExplainDisplay": "遗传算法像==方案相亲==：\n优胜者配对生娃，\n偶尔基因突变\n冒==黑马==。\n\n常用于路径、排班、调参，\n在巨大解空间里找较优解。", "relationsNarrative": "Optimization\nGA 是一种在大解空间里寻找更优解的优化方法。\n\nHeuristic Search\nGA 靠进化式试探搜索，不保证一步到最优。\n\nHPO\nGA 可用来自动试不同超参数组合。\n\nNo Free Lunch Theorem\n它提醒 GA 不会对所有问题都最好。", "relations": { "optimization": { "label": "用于求解…", "note": "它在巨大解空间里找更优方案。" }, "heuristic-search": { "label": "属于…", "note": "它靠经验式试探，而非穷举。" }, "hyperparameter-optimization": { "label": "可用于…", "note": "常被拿来自动挑参数组合。" }, "no-free-lunch-theorem": { "label": "受…约束", "note": "没有搜索法能通吃所有问题。" } } } } }, { "id": "gguf", "name": "GGUF", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "local-llm" }, { "to": "quantization" }, { "to": "open-weights" }, { "to": "vram" } ], "track": "today", "i18n": { "en": { "fullName": "GPT-Generated Unified Format", "factExplain": "A file format that helps load and run large AI models on your own computer.", "humanExplain": "GGUF is like giving a giant AI a carry-on bag. It is easier to move and fit on your laptop.\n\nYou see it when people share local models. It is common for quantized models you run offline.", "humanExplainDisplay": "GGUF is like giving a giant AI\n==a carry-on bag==.\nIt is easier to move\nand ==fit on your laptop==.\n\nYou see it when people share local models.\nIt is common for quantized models\nyou run offline.", "relationsNarrative": "Local-LLM\nGGUF is one common file format for running large models locally.\n\nQuantization\nGGUF often stores quantized models for lighter offline use.\n\nOpen weights\nOpen weight models are often converted to GGUF before sharing.\n\nVRAM\nGGUF often works with quantization to help models run with less VRAM.", "relations": { "local-llm": { "label": "is used by …", "note": "Local models often use GGUF files." }, "quantization": { "label": "often pairs with …", "note": "GGUF often stores quantized model weights." }, "open-weights": { "label": "packs … files", "note": "Open weights are often shared as GGUF files." }, "vram": { "label": "helps save …", "note": "GGUF plus quantization can lower VRAM pressure." } } }, "zh": { "fullName": "GPT-Generated Unified Format（模型统一文件格式）", "factExplain": "一种便于本地加载和运行大模型的文件格式。", "humanExplain": "它像把大模型塞进搬家真空袋，本来占一间屋，家用电脑也能扛上楼。\n\n它常配合量化和本地推理使用，方便下载、加载，也方便离线跑模型。", "humanExplainDisplay": "它像把大模型塞进==搬家真空袋==，\n本来占一间屋，\n==家用电脑也能扛上楼==。\n\n它常配合量化和本地推理使用，\n方便下载、加载，\n也方便离线跑模型。", "relationsNarrative": "Local-LLM\nGGUF 是本地部署大模型时最常见的文件格式之一。\n\nQuantization\nGGUF 常用来存放量化后的模型，方便轻量运行。\n\nOpen-weights\n开放权重模型常被转换成 GGUF 后再分发下载。\n\nVRAM\nGGUF 常配合量化，帮助模型在更小显存上运行。", "relations": { "local-llm": { "label": "常被…使用", "note": "本地运行模型时常见 GGUF 格式。" }, "quantization": { "label": "常与…搭配", "note": "GGUF 常承载量化后的模型权重。" }, "open-weights": { "label": "封装…文件", "note": "开放权重常以 GGUF 形式分发。" }, "vram": { "label": "帮忙省…", "note": "配合量化可降低显存占用压力。" } } } } }, { "id": "gibbs-sampling", "name": "Gibbs Sampling", "layer": "L2", "era": "1984", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "markov-chain-monte-carlo" }, { "to": "probabilistic-graphical-model" }, { "to": "bayesian-network" }, { "to": "boltzmann-machine" } ], "track": "history", "i18n": { "en": { "fullName": "Gibbs Sampling", "factExplain": "An MCMC method for sampling one variable at a time, using the others.", "humanExplain": "Gibbs Sampling is like a potluck group chat. It keeps the menu, then changes one sad salad after checking the other dishes.\n\nIt updates one unknown at a time in Bayesian models and graph models. Exact probability is too hard there.", "humanExplainDisplay": "Gibbs Sampling is like a ==potluck group chat==.\nIt keeps the menu,\nthen changes ==one sad salad==\nafter checking the other dishes.\n\nIt updates one unknown at a time\nin Bayesian models and graph models.\nExact probability is too hard there.", "relationsNarrative": "MCMC\nGibbs Sampling is an MCMC method using conditional samples to approach a target distribution.\n\nPGM\nIn a PGM, Gibbs Sampling estimates hard posterior probabilities.\n\nBayesian Network\nIn a Bayesian Network, Gibbs Sampling samples each variable from its conditional probability.\n\nBoltzmann Machine\nBoltzmann Machine training often uses Gibbs Sampling to draw energy-based samples.", "relations": { "markov-chain-monte-carlo": { "label": "belongs to …", "note": "It samples variables one by one and runs a Markov chain." }, "probabilistic-graphical-model": { "label": "does inference for …", "note": "PGMs often use it to estimate posterior probabilities." }, "bayesian-network": { "label": "samples in …", "note": "It works well with networks full of conditional links." }, "boltzmann-machine": { "label": "helps train …", "note": "It draws samples from the machine's energy distribution." } } }, "zh": { "fullName": "吉布斯采样", "factExplain": "一种逐个变量条件采样的 MCMC 方法。", "humanExplain": "吉布斯采样像家庭群接龙：不全员重填，先看别人答案，再改自己那格。\n\n用于贝叶斯和图模型，逼近算不动的概率。", "humanExplainDisplay": "吉布斯采样像家庭群接龙：\n不全员重填，\n先看==别人答案==，\n再改==自己那格==。\n\n用于贝叶斯和图模型，\n逼近算不动的概率。", "relationsNarrative": "MCMC\nGibbs Sampling 是 MCMC 的一种，靠条件采样逼近目标分布。\n\nPGM\n在图模型里，它常用来近似难以直接计算的后验。\n\nBayesian Network\n贝叶斯网络可用它按条件概率轮流抽样。\n\nBoltzmann Machine\n玻尔兹曼机训练常用它从能量分布中取样。", "relations": { "markov-chain-monte-carlo": { "label": "属于…方法", "note": "逐个变量采样，跑出马尔可夫链。" }, "probabilistic-graphical-model": { "label": "为…做推断", "note": "图模型常用它近似后验分布。" }, "bayesian-network": { "label": "在…中采样", "note": "适合处理条件依赖复杂的网络。" }, "boltzmann-machine": { "label": "训练…时常用", "note": "常用它在能量模型里取样。" } } } } }, { "id": "glm-5-2", "name": "GLM-5.2", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "multimodal" }, { "to": "foundation-model" }, { "to": "qwen" } ], "track": "today", "i18n": { "en": { "fullName": "Zhipu GLM-5.2 Model", "factExplain": "A large language model version released by Zhipu AI.", "humanExplain": "GLM-5.2 is like a new do-it-all counter at the mall food court. Say one thing, and it starts cooking with words.\n\nIt can write, answer questions, and sometimes handle images. Use it as a general helper or as the base for business apps.", "humanExplainDisplay": "GLM-5.2 is like a new ==do-it-all counter==\nat the mall food court.\nSay one thing,\nand it starts ==cooking with words==.\n\nIt can write, answer questions,\nand sometimes handle images.\nUse it as a general helper\nor as the base for business apps.", "relationsNarrative": "LLM\nGLM-5.2 is part of the LLM family.\n\nMultimodal AI\nIf it can use text and images, it has multimodal skills.\n\nFoundation-model\nGLM-5.2 can be a base model for tuning or apps.\n\nQwen\nPeople often compare GLM-5.2 with Qwen for fit and skill.", "relations": { "llm": { "label": "is a kind of …", "note": "It is a large language model at heart." }, "multimodal": { "label": "can have … skills", "note": "If it can use images, it follows the multimodal path." }, "foundation-model": { "label": "can serve as a …", "note": "Teams can fine-tune it or plug it into apps." }, "qwen": { "label": "is often compared with …", "note": "Both are common model choices in the Chinese AI world." } } }, "zh": { "fullName": "智谱 GLM-5.2 模型", "factExplain": "智谱推出的一代大语言模型。", "humanExplain": "像店里新添了个万能档口：顾客一句话点单，它能写、能答、还能看图上菜。\n\n适合做通用助手，也可接入企业系统当模型底座。", "humanExplainDisplay": "像店里新添了个\n==万能档口==：\n顾客一句话点单，\n它能写、能答、还能==看图上菜==。\n\n适合做通用助手，\n也可接入企业系统，\n当模型底座。", "relationsNarrative": "LLM\n它属于大语言模型家族，是通用生成模型的一员。\n\nMultimodal\n若支持图文输入输出，它就具备多模态能力。\n\nFoundation-model\n它可作为基础模型，被继续微调或接入应用。\n\nQwen\n它常与 Qwen 一起被拿来比较能力与场景适配。", "relations": { "llm": { "label": "属于…一类", "note": "它本质上是一种大语言模型。" }, "multimodal": { "label": "可具备…能力", "note": "若支持看图，就属于多模态路线。" }, "foundation-model": { "label": "可作为…底座", "note": "能被继续微调或接入应用。" }, "qwen": { "label": "与…同台比较", "note": "都属于中文世界常见模型选择。" } } } } }, { "id": "global-workspace-theory", "name": "GWT", "layer": "L1", "era": "1988", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "ai-consciousness-debate" }, { "to": "cognitive-model" }, { "to": "attention" }, { "to": "agi" } ], "track": "today", "seo": { "zh": { "title": "全局工作空间理论 是什么?脑内的导播台,一文看懂 — AI Rookies", "description": "解释意识如何向全脑广播信息的认知理论,也是讨论 AI 有没有意识的常用框架。人话版一分钟讲明白。" }, "en": { "title": "What Is Global Workspace Theory? The Brain's PA System", "description": "The consciousness theory behind today's AI debates: one thought gets the microphone, the whole brain hears it. Explained simply, with a concept map." } }, "i18n": { "en": { "fullName": "Global Workspace Theory", "factExplain": "A theory explaining consciousness as information broadcast across the mind.", "humanExplain": "GWT is the school PA system in your head. Once a thought gets the microphone, the whole brain hears it.\n\nIt helps people talk about machine consciousness. It also inspires AI with many parts working together.", "humanExplainDisplay": "GWT is the ==school PA system== in your head.\nOnce a thought gets ==the microphone==,\nthe whole brain hears it.\n\nIt helps people talk about machine consciousness.\nIt also inspires AI with many parts working together.", "relationsNarrative": "AI Consciousness\nGWT is a common frame for debates about AI consciousness.\n\nCognitive Model\nGWT is a Cognitive Model about information broadcast in the mind.\n\nAttention\nAttention works like the gate into the global workspace.\n\nAGI\nSome AGI designs borrow its idea of a global broadcast.", "relations": { "ai-consciousness-debate": { "label": "offers a frame for …", "note": "People use GWT in debates about AI consciousness." }, "cognitive-model": { "label": "is a kind of …", "note": "It treats consciousness as information broadcast." }, "attention": { "label": "filters through …", "note": "Attention acts like the tryout before the main stage." }, "agi": { "label": "inspires … designs", "note": "Some AGI ideas borrow its global broadcast." } } }, "zh": { "fullName": "全局工作空间理论（Global Workspace Theory）", "factExplain": "一种解释意识如何广播信息的认知理论。", "humanExplain": "GWT 就是春晚导播台：节目切上主屏，全脑观众才一起看见。\n\n用于解释机器意识，也启发多模块 AI 协作。", "humanExplainDisplay": "GWT 就是==春晚导播台==：\n节目==切上主屏==，\n全脑观众才一起看见。\n\n用于解释机器意识，\n也启发多模块 AI 协作。", "relationsNarrative": "AI Consciousness\nGWT 是判断 AI 是否可能有意识的常用参照。\n\nCognitive Model\n它属于认知模型，用广播解释意识整合。\n\nAttention\n注意力常被视为进入全局空间的筛选门。\n\nAGI\n一些 AGI 架构借鉴它来组织模块协作。", "relations": { "ai-consciousness-debate": { "label": "为…提供框架", "note": "常被用来讨论 AI 是否有意识。" }, "cognitive-model": { "label": "属于…的一种", "note": "它把意识解释成信息广播机制。" }, "attention": { "label": "借…筛选信息", "note": "注意力像进入舞台前的选拔。" }, "agi": { "label": "启发…架构", "note": "一些 AGI 设想借鉴全局广播。" } } } } }, { "id": "glove", "name": "GloVe", "layer": "L3", "era": "2014", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "embedding" }, { "to": "word2vec" }, { "to": "distributional-semantics" }, { "to": "matrix-factorization" } ], "track": "history", "i18n": { "en": { "fullName": "Global Vectors for Word Representation", "factExplain": "GloVe learns word vectors from how often words appear near other words.", "humanExplain": "GloVe is like watching the school cafeteria for a month. Words with the same lunch crowd get seats near each other.\n\nIt makes fixed word embeddings. People use them for search. They also help sort text and match meanings.", "humanExplainDisplay": "GloVe is like watching the ==school cafeteria== for a month.\nWords with the same ==lunch crowd== get seats near each other.\n\nIt makes fixed word embeddings.\nPeople use them for search.\nThey also help sort text and match meanings.", "relationsNarrative": "Embedding\nGloVe's main output is a word embedding a computer can compare.\n\nWord2Vec\nGloVe and Word2Vec are both classic fixed word embedding methods.\n\nDist. Semantics\nGloVe turns the idea of learning meaning from word neighbors into a trainable method.\n\nMatrix Factorization\nGloVe is basically factorizing a table of word co-occurrences.", "relations": { "embedding": { "label": "produces …", "note": "GloVe turns words into number positions computers can compare." }, "word2vec": { "label": "contrasts with …", "note": "Both are classic ways to make fixed word embeddings." }, "distributional-semantics": { "label": "builds on …", "note": "A word's meaning comes from the words around it." }, "matrix-factorization": { "label": "uses …", "note": "GloVe can be seen as splitting a word co-occurrence table." } } }, "zh": { "fullName": "Global Vectors for Word Representation，全局词向量表示", "factExplain": "利用全局共现统计学习词向量的方法。", "humanExplain": "GloVe像小区大妈认邻居：总一起遛弯的词，语义就住得近。\n\n产出静态词向量，常用于搜索、分类和语义匹配。", "humanExplainDisplay": "GloVe像小区大妈认邻居：\n总一起==遛弯的词==，\n语义就==住得近==。\n\n产出静态词向量，\n常用于搜索、分类\n和语义匹配。", "relationsNarrative": "Embedding\nGloVe 的直接产物，就是可计算的词向量。\n\nWord2Vec\n它和 Word2Vec 都是经典静态词向量。\n\nDistributional Semantics\n它把“看邻居猜词义”做成可训练方法。\n\nMatrix Factorization\n它本质上在分解词与词的共现矩阵。", "relations": { "embedding": { "label": "产出…表示", "note": "它把词变成可计算的向量坐标。" }, "word2vec": { "label": "对比…方法", "note": "两者都是经典静态词向量方法。" }, "distributional-semantics": { "label": "继承…思想", "note": "词义来自它和哪些词一起出现。" }, "matrix-factorization": { "label": "借用…分解", "note": "它可看作分解词共现矩阵。" } } } } }, { "id": "government-ai-assistant", "name": "Government AI Assistant", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "rag" }, { "to": "hallucination" }, { "to": "data-privacy" }, { "to": "ai-governance-framework" } ], "track": "today", "i18n": { "en": { "fullName": "Government AI Assistant", "factExplain": "An AI helper built for government services and office work.", "humanExplain": "Think of the city hall clerk who knows every form. Ask about a permit, and they point to the right window before lunch ends.\n\nIt helps answer policy questions and guide public services. A wrong answer can send people down the wrong line.", "humanExplainDisplay": "Think of the ==city hall clerk==\nwho knows every form.\nAsk about a permit,\nand they point to the ==right window==\nbefore lunch ends.\n\nIt helps answer policy questions\nand guide public services.\nA wrong answer can send people\ndown the wrong line.", "relationsNarrative": "RAG\nIt often uses RAG to find policy files before it answers.\n\nHallucination\nA hallucination can mislead people about rules or steps.\n\nData-privacy\nIt may see IDs and private records, so privacy matters more.\n\nAI Governance\nAI Governance sets clear limits and responsibility for its use.", "relations": { "rag": { "label": "looks up policy with …", "note": "Government Q&A often uses RAG to find the newest files." }, "hallucination": { "label": "must prevent …", "note": "One made-up policy line can send people the wrong way." }, "data-privacy": { "label": "must protect …", "note": "It may handle IDs and private records." }, "ai-governance-framework": { "label": "works within …", "note": "Rules set what it can do and who is responsible." } } }, "zh": { "fullName": "政务 AI 助手", "factExplain": "面向政府服务与办公场景的 AI 助手。", "humanExplain": "像带团的金牌导游，几十处窗口、要带的材料、容易卡的环节全门儿清，你只管跟着走。\n\n常用于政策问答和办事导引，答错政策就可能误事。", "humanExplainDisplay": "像带团的==金牌导游==，\n几十处窗口、要带的材料、\n容易卡的环节全门儿清，\n你==只管跟着走==。\n\n常用于政策问答\n和办事导引，\n答错政策就可能误事。", "relationsNarrative": "RAG\n它常靠 RAG 检索政策和办事资料，再组织回答。\n\nHallucination\n一旦出现幻觉，政务答复就可能误导公众或流程。\n\nData-privacy\n它常接触身份、材料等敏感信息，隐私要求更高。\n\nAI Governance\n政务场景通常要在治理框架下明确边界与责任。", "relations": { "rag": { "label": "常配合…查政策", "note": "政务问答常靠检索最新文件。" }, "hallucination": { "label": "需要压住…", "note": "政策场景里编错一句都可能误事。" }, "data-privacy": { "label": "必须处理…风险", "note": "常接触个人与敏感政务信息。" }, "ai-governance-framework": { "label": "落在…约束内", "note": "上线往往受制度与责任边界约束。" } } } } }, { "id": "gpt-3", "name": "GPT-3", "layer": "L3", "era": "2020", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "gpt" }, { "to": "llm" }, { "to": "scaling-law" }, { "to": "emergence" } ], "track": "history", "i18n": { "en": { "fullName": "Generative Pre-trained Transformer 3", "factExplain": "A large text-generating AI model released by OpenAI in 2020.", "humanExplain": "GPT-3 felt like the new kid in the cafeteria. Somehow, it could chat at every table.\n\nIt made AI chat and writing feel real to many people. It helped push LLMs into the spotlight.", "humanExplainDisplay": "GPT-3 felt like the ==new kid in the cafeteria==.\nSomehow, it could ==chat at every table==.\n\nIt made AI chat and writing feel real to many people.\nIt helped push LLMs into the spotlight.", "relationsNarrative": "GPT\nGPT-3 is a key GPT generation. It follows the same bigger-is-stronger path.\n\nLLM\nGPT-3 helped move LLMs from expert talk into everyday talk.\n\nScaling-law\nGPT-3 became a famous example of gaining power by scaling up.\n\nEmergence\nGPT-3 showed how new skills can appear when a model gets bigger.", "relations": { "gpt": { "label": "belongs to … family", "note": "GPT-3 is a key generation in the GPT line." }, "llm": { "label": "brought … into public view", "note": "GPT-3 made LLMs something many people noticed." }, "scaling-law": { "label": "showed the … idea", "note": "It showed how bigger models can gain stronger skills." }, "emergence": { "label": "showed …", "note": "New skills appeared after the model got very large." } } }, "zh": { "fullName": "生成式预训练 Transformer 3", "factExplain": "OpenAI 于 2020 年发布的大规模生成式语言模型。", "humanExplain": "那会儿它一张嘴，像班里突然来了个什么都能聊两句的转校生，大家才发现 AI 真能接话。\n\n它把大模型聊天写作带出圈，让更多人第一次认真关注 LLM。", "humanExplainDisplay": "那会儿它一张嘴，\n像班里突然来了个\n什么都能聊两句的==转校生==，\n大家才发现 AI\n真能==接话==。\n\n它把大模型聊天写作带出圈，\n让更多人第一次\n认真关注 LLM。", "relationsNarrative": "Gpt\n它是 GPT 系列中的代表性一代，延续同一路线放大能力。\n\nLlm\n它让大语言模型从圈内概念，变成大众都听过的东西。\n\nScaling-law\n它常被视为“堆规模能变强”这套思路的代表案例。\n\nEmergence\n它让人们更直观看到，模型变大后会冒出新能力。", "relations": { "gpt": { "label": "属于…家族", "note": "它是 GPT 系列里关键一代。" }, "llm": { "label": "推动…出圈", "note": "它让大语言模型进入大众视野。" }, "scaling-law": { "label": "体现…思路", "note": "靠更大规模换来更强能力。" }, "emergence": { "label": "展现…现象", "note": "规模上去后冒出意外能力。" } } } } }, { "id": "gpt-5-5", "name": "GPT-5.5", "layer": "L3", "era": "2026", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "gpt" }, { "to": "llm" }, { "to": "chatgpt" }, { "to": "frontier-model" } ], "track": "history", "i18n": { "en": { "fullName": "OpenAI GPT-5.5", "factExplain": "A large language model version released between major generations.", "humanExplain": "GPT-5.5 is like a video game patch. No new castle, but the lag and weird buttons feel less annoying.\n\nYou may meet it in chat helpers and coding tools. It is mostly polish, not a whole new AI idea.", "humanExplainDisplay": "GPT-5.5 is like a ==video game patch==.\nNo new castle,\nbut the ==lag and weird buttons== feel less annoying.\n\nYou may meet it in chat helpers and coding tools.\nIt is mostly polish,\nnot a whole new AI idea.", "relationsNarrative": "GPT\nGPT-5.5 belongs to the GPT family as a between-generation version.\n\nLLM\nGPT-5.5 is still an LLM, just with a finer version name.\n\nChatGPT\nModels like GPT-5.5 often power chat products and shape daily use.\n\nFrontier model\nIf GPT-5.5 leads the field, people may call it a frontier model.", "relations": { "gpt": { "label": "belongs to …", "note": "It is one version step in the GPT family." }, "llm": { "label": "is a kind of …", "note": "It is still a large language model at heart." }, "chatgpt": { "label": "often powers …", "note": "Models like this often power chat products." }, "frontier-model": { "label": "may count as …", "note": "If it leads the field, people call it a frontier model." } } }, "zh": { "fullName": "OpenAI 第 5.5 代生成式预训练模型", "factExplain": "一个介于代际之间发布的大语言模型版本。", "humanExplain": "它更像手机系统的小版本更新：界面没大换血，但卡顿、续航、手势这些都被==悄悄修顺==，用久了就觉得==没那么别扭==。\n\n常见于聊天助手和开发工具，重点是把体验打磨得更稳，不一定代表范式大变。", "humanExplainDisplay": "它更像手机系统的小版本更新：\n界面没大换血，\n但卡顿、续航、手势这些都被==悄悄修顺==，\n用久了就觉得==没那么别扭==。\n\n常见于聊天助手和开发工具，\n重点是把体验打磨得更稳，\n不一定代表范式大变。", "relationsNarrative": "GPT\n它属于 GPT 系列，是一次非整代命名的版本迭代。\n\nLLM\n它本质上仍是大语言模型，只是代际定位更细。\n\nChatGPT\n这类模型常被接入聊天产品，直接影响日常使用体验。\n\nFrontier model\n若能力处于行业前列，它通常会被视作前沿模型。", "relations": { "gpt": { "label": "属于…系列", "note": "它是 GPT 系列中的一次版本迭代。" }, "llm": { "label": "是一种…", "note": "本质上仍是大语言模型。" }, "chatgpt": { "label": "常被用于…", "note": "这类模型常作为聊天产品底层能力。" }, "frontier-model": { "label": "常被视作…", "note": "若能力领先，通常会被归入前沿模型。" } } } } }, { "id": "gpt-5-6-sol", "name": "GPT-5.6 Sol", "layer": "L3", "era": "2026", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "gpt-5-5" }, { "to": "gpt" }, { "to": "llm" }, { "to": "frontier-model" } ], "track": "today", "i18n": { "en": { "fullName": "OpenAI GPT-5.6 Sol", "factExplain": "A high-end general AI model in OpenAI’s GPT-5.6 series.", "humanExplain": "GPT-5.6 Sol is a driver with a new engine. Same road, but faster and smoother. It even tells you why it took each turn.\n\nUse it for chat, coding, and hard reasoning. It gets you to a good plan faster.", "humanExplainDisplay": "GPT-5.6 Sol is a driver\nwith a ==new engine==.\nSame road, but faster and smoother.\nIt even tells you ==why it took each turn==.\n\nUse it for chat, coding, and hard reasoning.\nIt gets you to a good plan faster.", "relationsNarrative": "GPT-5.5\nGPT-5.6 Sol is a later step in the GPT-5.x line.\n\nGPT\nGPT-5.6 Sol continues the GPT family path for general language models.\n\nLLM\nGPT-5.6 Sol is still a kind of large language model.\n\nFrontier model\nGPT-5.6 Sol is part of the race for the strongest general models.", "relations": { "gpt-5-5": { "label": "follows …", "note": "It is a later GPT-5.x model after GPT-5.5." }, "gpt": { "label": "continues the … family", "note": "It stays on the GPT path for general language models." }, "llm": { "label": "is a kind of …", "note": "At its core, it is still a large language model." }, "frontier-model": { "label": "competes as a …", "note": "It is part of the race among the strongest general models." } } }, "zh": { "fullName": "OpenAI GPT-5.6 Sol 模型", "factExplain": "OpenAI GPT-5.6 系列的高端通用大模型。", "humanExplain": "GPT-5.6 Sol 像老司机换了新引擎：同一条路开得更快更稳，还顺嘴讲清每个弯为啥这么拐。\n\n适合聊天、编程和硬核推理，出方案更快。", "humanExplainDisplay": "GPT-5.6 Sol 像老司机\n==换了新引擎==：\n同一条路更快更稳，\n还顺嘴讲清==每个弯为啥这么拐==。\n\n适合聊天、编程和硬核推理，\n出方案更快。", "relationsNarrative": "GPT-5.5\n它是 GPT-5.x 路线上的后续迭代型号。\n\nGPT\n它延续 GPT 家族的通用语言模型路线。\n\nLLM\n它本质上仍是大语言模型的一种。\n\nFrontier model\n它参与前沿模型的能力竞赛。", "relations": { "gpt-5-5": { "label": "接替…迭代", "note": "同属 GPT-5.x 的后续型号。" }, "gpt": { "label": "延续…家族", "note": "沿用 GPT 系列通用模型路线。" }, "llm": { "label": "属于…", "note": "本质仍是大语言模型。" }, "frontier-model": { "label": "冲击…", "note": "面向最强通用能力竞争。" } } } } }, { "id": "gpt-5-6", "name": "GPT-5.6", "layer": "L3", "era": "2026", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "gpt" }, { "to": "gpt-5-5" }, { "to": "gpt-5-6-sol" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "OpenAI GPT-5.6 Model", "factExplain": "A general language model version in OpenAI’s GPT-5 family.", "humanExplain": "GPT-5.6 is like an overnight phone update. Same icon. Less lag. Fewer weird pop-ups.\n\nYou meet it in coding and office work. It also helps with hard Q&A. Its reasoning is steadier. Its tools fumble less.", "humanExplainDisplay": "GPT-5.6 is like an ==overnight phone update==.\nSame icon.\nLess lag.\nFewer ==weird pop-ups==.\n\nYou meet it in coding and office work.\nIt also helps with hard Q&A.\nIts reasoning is steadier.\nIts tools fumble less.", "relationsNarrative": "GPT\nGPT-5.6 continues the general model path of the GPT family.\n\nGPT-5.5\nGPT-5.6 is usually seen as the next step after GPT-5.5.\n\nGPT-5.6 Sol\nGPT-5.6 Sol is more like a focused version from the same generation.\n\nLLM\nGPT-5.6 is still an LLM built for language tasks.", "relations": { "gpt": { "label": "continues … line", "note": "It is an updated model in the GPT family." }, "gpt-5-5": { "label": "iterates from …", "note": "GPT-5.6 is usually seen as the next step after GPT-5.5." }, "gpt-5-6-sol": { "label": "sits beside …", "note": "Sol is a same-generation version with a more focused role." }, "llm": { "label": "is a type of …", "note": "It is still a large language model at heart." } } }, "zh": { "fullName": "OpenAI GPT-5.6 模型", "factExplain": "OpenAI GPT-5 系列的通用语言模型版本。", "humanExplain": "GPT-5.6像手机半夜偷升级：图标没换，打开更顺，弹窗少作妖。\n\n常用于代码、办公和复杂问答，推理更稳，工具更少掉链。", "humanExplainDisplay": "GPT-5.6像手机\n==半夜偷升级==：\n图标没换，打开更顺，\n弹窗==少作妖==。\n\n常用于代码、办公\n和复杂问答，\n推理更稳，工具更少掉链。", "relationsNarrative": "GPT\n它延续 GPT 家族的通用生成模型路线。\n\nGPT-5.5\nGPT-5.6 通常被看作 GPT-5.5 的后续迭代。\n\nGPT-5.6 Sol\nSol 更像同代里的专项或强化版本。\n\nLLM\n它本质上仍是面向语言任务的大语言模型。", "relations": { "gpt": { "label": "继承…路线", "note": "它属于 GPT 模型家族的一代更新。" }, "gpt-5-5": { "label": "迭代自…", "note": "5.6 通常被看作 5.5 的后续版本。" }, "gpt-5-6-sol": { "label": "并列于…", "note": "Sol 可视为同代的专项版本。" }, "llm": { "label": "属于…", "note": "它本质上仍是大语言模型。" } } } } }, { "id": "gpt-live", "name": "GPT-Live", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "gpt" }, { "to": "streaming-multimodal-model" }, { "to": "voice-to-voice-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Real-time Multimodal GPT", "factExplain": "A GPT model built for live voice and camera interaction.", "humanExplain": "GPT-Live is like FaceTiming an AI friend. You pause mid-sentence, and it spots the dog stealing toast.\n\nYou meet it in customer support and practice coaches. It can also guide you on video. Low delay keeps the chat from freezing.", "humanExplainDisplay": "GPT-Live is like ==FaceTiming an AI friend==.\nYou pause mid-sentence,\nand it spots ==the dog stealing toast==.\n\nYou meet it in customer support\nand practice coaches.\nIt can also guide you on video.\nLow delay keeps the chat\nfrom freezing.", "relationsNarrative": "GPT\nGPT-Live is the GPT form built for live interaction.\n\nLive Multimodal\nGPT-Live uses Live Multimodal to listen, watch, and respond as things happen.\n\nVoice-to-voice-ai\nGPT-Live supports Voice-to-voice-ai for phone-like speech chats.", "relations": { "gpt": { "label": "belongs to … family", "note": "It is the GPT form built for live interaction." }, "streaming-multimodal-model": { "label": "uses … style", "note": "It listens, watches, and responds as things happen." }, "voice-to-voice-ai": { "label": "supports … chats", "note": "It makes spoken AI feel more like a phone call." } } }, "zh": { "fullName": "实时多模态 GPT", "factExplain": "面向实时语音、视觉交互的 GPT 模型。", "humanExplain": "GPT-Live像把AI塞进视频通话：你话音未落，它能接梗，还会瞄一眼现场。\n\n用于实时客服、陪练、视频助手，低延迟让对话不断片。", "humanExplainDisplay": "GPT-Live像把AI塞进\n==视频通话==：\n你话音未落，它能接梗，\n还会==瞄一眼现场==。\n\n用于实时客服、陪练、视频助手，\n低延迟让对话不断片。", "relationsNarrative": "GPT\n它是 GPT 家族面向实时交互的形态。\n\nLive Multimodal\n它采用实时多模态形态，边听边看边生成。\n\nVoice-to-voice AI\n它支持端到端语音对话，体验更像打电话。", "relations": { "gpt": { "label": "属于…家族", "note": "它是 GPT 面向实时交互的形态。" }, "streaming-multimodal-model": { "label": "采用…形态", "note": "边听边看边生成，减少等待感。" }, "voice-to-voice-ai": { "label": "支持…体验", "note": "让语音对话更像电话聊天。" } } } } }, { "id": "gpt", "name": "GPT", "layer": "L3", "era": "2018", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "pretraining" }, { "to": "gpt-3" }, { "to": "llm" } ], "track": "history", "i18n": { "en": { "fullName": "Generative Pre-trained Transformer", "factExplain": "A Transformer language model pre-trained first, then used to generate text.", "humanExplain": "GPT is like a kid after a summer in the library. Give it “Once upon a time,” and it keeps the story train going.\n\nYou meet it in chatbots and writing tools. It also powers code helpers and many big AI products.", "humanExplainDisplay": "GPT is like a kid after a ==summer in the library==.\nGive it “Once upon a time,”\nand it ==keeps the story train going==.\n\nYou meet it in chatbots and writing tools.\nIt also powers code helpers\nand many big AI products.", "relationsNarrative": "Transformer\nGPT is built on Transformer and became a main branch of that design.\n\nPretraining\nGPT starts with large-scale Pretraining, then gains general text skills.\n\nGPT-3\nGPT-3 brought GPT into public view and made it a hot topic.\n\nLLM\nGPT is a classic member of the LLM family.", "relations": { "transformer": { "label": "is built on …", "note": "GPT is a branch of the Transformer design." }, "pretraining": { "label": "starts with …", "note": "GPT learns from huge text sets before it does general language tasks." }, "gpt-3": { "label": "broke out with …", "note": "GPT-3 made this model family famous around the world." }, "llm": { "label": "is a classic …", "note": "GPT is one of the best-known kinds of LLM." } } }, "zh": { "fullName": "生成式预训练 Transformer", "factExplain": "一种先预训练、再生成文本的 Transformer 语言模型。", "humanExplain": "先把全网文字闷头啃一大遍，等你开口时，它就像故事接龙王，顺着你的话往下续。\n\n常做聊天、写作和代码补全，是很多大模型产品的基础路线。", "humanExplainDisplay": "先把全网文字闷头\n==啃一大遍==，\n等你开口时，\n它就像==故事接龙王==，\n顺着你的话往下续。\n\n常做聊天、\n写作和代码补全，\n是很多大模型产品的\n基础路线。", "relationsNarrative": "Transformer\n它建立在 Transformer 架构上，是这条技术路线的代表分支。\n\nPretraining\n它先靠大规模预训练打底，再获得通用的语言生成能力。\n\nGpt-3\nGPT-3 把这类模型推到大众视野，成了行业爆点。\n\nLLM\n它是大语言模型家族里的典型代表，很多人提到大模型先想到它。", "relations": { "transformer": { "label": "建立在…之上", "note": "它本质上是 Transformer 架构的一支。" }, "pretraining": { "label": "先做…训练", "note": "先大规模预训练，再具备通用语言能力。" }, "gpt-3": { "label": "在…中出圈", "note": "GPT-3 让这条路线真正火遍全球。" }, "llm": { "label": "属于…代表", "note": "它是大语言模型最典型的一类。" } } } } }, { "id": "gpu", "name": "GPU", "layer": "L5", "sublayer": "compute", "era": "2012", "publishedAt": "2026-05-23T10:15:00Z", "relations": [ { "to": "pretraining" }, { "to": "inference" }, { "to": "compute-race" }, { "to": "foundation-model" } ], "i18n": { "en": { "fullName": "GPU Compute", "factExplain": "Hardware built to run many calculations at once, powering AI training and inference.", "humanExplain": "A GPU is like a cafeteria kitchen at lunch rush. Not chatty, but it flips a thousand pancakes at once.\n\nIt trains big models. It runs AI replies. It helps make images. That is why GPU time can cost so much.", "humanExplainDisplay": "A GPU is like a ==cafeteria kitchen==\nat lunch rush.\nNot chatty,\nbut it flips ==a thousand pancakes at once==.\n\nIt trains big models.\nIt runs AI replies.\nIt helps make images.\nThat is why GPU time can cost so much.", "relationsNarrative": "Pretraining\nPretraining depends more on GPU clusters as it gets larger.\n\nInference\nGPU power often decides Inference speed and capacity.\n\nCompute-race\nTight GPU supply can make the Compute-race more intense.\n\nFoundation-model\nA Foundation-model often needs many GPUs for a long training run.", "relations": { "pretraining": { "label": "powers …", "note": "Bigger Pretraining needs more GPU clusters for longer." }, "inference": { "label": "runs …", "note": "GPU power often decides how fast Inference feels." }, "compute-race": { "label": "fuels …", "note": "Tight GPU supply can make the Compute-race hotter." }, "foundation-model": { "label": "trains …", "note": "A Foundation-model can occupy many GPUs for a long time." } } }, "zh": { "fullName": "GPU 算力", "factExplain": "适合并行计算、支撑 AI 训练和推理的核心硬件。", "humanExplain": "它像后厨一排切菜阿姨：不负责想菜单，只负责同时剁一万根葱。\n\n它常用于训练和推理，决定模型能跑多快、能跑多大。", "humanExplainDisplay": "它像==后厨一排切菜阿姨==：\n不负责想菜单，\n只负责==同时剁一万根葱==。\n\n它常用于训练和推理，\n决定模型能跑多快、能跑多大。", "relationsNarrative": "Pretraining\nPretraining 规模越大，对 GPU 集群的依赖越强。\n\nInference\nInference 延迟和吞吐，往往由 GPU 能力决定。\n\nCompute-race\nGPU 供应越紧张，Compute-race 就越容易加剧。\n\nFoundation-model\nFoundation-model 的训练通常需要长期占用大量 GPU。", "relations": { "pretraining": { "label": "支撑…" }, "inference": { "label": "支撑…运行" }, "compute-race": { "label": "引发…" }, "foundation-model": { "label": "训练…的算力" } } } } }, { "id": "gradient-boosting", "name": "Gradient Boosting", "layer": "L2", "era": "1999", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "gradient-descent" }, { "to": "classification" }, { "to": "regression" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Gradient Boosting", "factExplain": "An ensemble method that adds weak models to fix earlier mistakes.", "humanExplain": "Gradient Boosting is like a team of tiny tutors. Each new tutor works on yesterday’s wrong answers.\n\nYou meet it in spreadsheets for yes-or-no calls and number guesses. It keeps fixing earlier errors, round by round.", "humanExplainDisplay": "Gradient Boosting is like\na team of ==tiny tutors==.\nEach new tutor works on\n==yesterday’s wrong answers==.\n\nYou meet it in spreadsheets\nfor yes-or-no calls and number guesses.\nIt keeps fixing earlier errors,\nround by round.", "relationsNarrative": "Gradient Descent\nGradient Boosting uses Gradient Descent to choose the next fix.\n\nClassification\nGradient Boosting is often used for spam checks and risk decisions.\n\nRegression\nGradient Boosting also predicts numbers, like home prices or sales.\n\nBias-Variance Tradeoff\nToo many deep trees can lower bias but raise variance.", "relations": { "gradient-descent": { "label": "uses … to steer fixes", "note": "Each round moves downhill on the loss score." }, "classification": { "label": "works well for …", "note": "It is a classic tool for sorting cases into groups." }, "regression": { "label": "also handles …", "note": "It is strong at predicting numbers like price or sales." }, "bias-variance-tradeoff": { "label": "shifts the …", "note": "Too many rounds can fit the training data too tightly." } } }, "zh": { "fullName": "Gradient Boosting｜梯度提升", "factExplain": "逐轮拟合残差并叠加弱模型的集成学习方法。", "humanExplain": "梯度提升像补作业：前一题错哪儿，后一轮就盯着那道补，越补越像标准答案。\n\n常用于表格分类和回归，能持续修正前几轮错误。", "humanExplainDisplay": "梯度提升像==补作业==：\n前一题错哪儿，\n后一轮就==盯着那道补==，\n越补越像标准答案。\n\n常用于表格分类和回归，\n能持续修正前几轮错误。", "relationsNarrative": "Gradient Descent\n它借损失函数的梯度，决定每一轮该往哪儿补错。\n\nClassification\n它常被用于垃圾邮件识别、风控审批这类分类任务。\n\nRegression\n它也常做房价、销量这类连续数值预测。\n\nBias-Variance Tradeoff\n树太多或太深时，可能压低偏差却抬高方差。", "relations": { "gradient-descent": { "label": "借…来纠错", "note": "每轮沿损失下降方向继续补错。" }, "classification": { "label": "常用于…任务", "note": "它是经典分类模型家族成员。" }, "regression": { "label": "也用于…预测", "note": "连续数值预测是它的拿手戏。" }, "bias-variance-tradeoff": { "label": "影响…平衡", "note": "轮数过多时更容易学得太死。" } } } } }, { "id": "gradient-clipping", "name": "Gradient Clipping", "layer": "L2", "era": "2013", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "backpropagation" }, { "to": "gradient-descent" }, { "to": "sgd" }, { "to": "lstm" } ], "track": "history", "i18n": { "en": { "fullName": "Gradient Clipping", "factExplain": "A technique that caps huge gradients so training updates do not fly out of control.", "humanExplain": "Gradient clipping is a speed limiter on a go-kart. The kid stomps the gas, but the snacks survive.\n\nYou meet it in RNN and LLM training. It tames exploding gradients, so training steps stay steady.", "humanExplainDisplay": "Gradient clipping is a ==speed limiter== on a go-kart.\nThe kid stomps the gas,\nbut the ==snacks survive==.\n\nYou meet it in RNN and LLM training.\nIt tames exploding gradients,\nso training steps stay steady.", "relationsNarrative": "Backpropagation\nBackpropagation finds the gradients, and clipping blocks the huge ones.\n\nGradient Descent\nGradient Descent updates parameters, and clipping stops steps from getting too big.\n\nSGD\nSGD can make noisy updates, and clipping helps keep training under control.\n\nLSTM\nLSTM models often use clipping to reduce exploding gradients.", "relations": { "backpropagation": { "label": "clips huge gradients from …", "note": "Backpropagation can produce gradients that suddenly get too large." }, "gradient-descent": { "label": "limits steps in …", "note": "It sets a cap on each parameter update." }, "sgd": { "label": "steadies … training", "note": "When SGD gradients jump too high, clipping can cut them down." }, "lstm": { "label": "protects … training", "note": "LSTM training often faces exploding gradients." } } }, "zh": { "fullName": "梯度裁剪", "factExplain": "限制梯度大小，防止训练更新失控。", "humanExplain": "梯度裁剪像高压锅限压阀：火再旺也先放气，别让参数把锅盖顶飞。\n\n用于 RNN、LLM 训练，压住梯度爆炸，让优化更稳。", "humanExplainDisplay": "梯度裁剪像==高压锅限压阀==：\n火再旺也先放气，\n别让参数\n把==锅盖顶飞==。\n\n用于 RNN、LLM 训练，\n压住梯度爆炸，\n让优化更稳。", "relationsNarrative": "Backpropagation\n反向传播算出梯度，裁剪负责拦住异常大值。\n\nGradient Descent\n梯度下降按梯度更新参数，裁剪限制步子过大。\n\nSGD\nSGD 更新噪声大时，裁剪能减少训练失控。\n\nLSTM\nLSTM 这类循环模型常用它缓解梯度爆炸。", "relations": { "backpropagation": { "label": "裁掉…中的过大梯度", "note": "反向传播算出的梯度可能突然爆表。" }, "gradient-descent": { "label": "限制…的步子", "note": "它给每次参数更新设上限。" }, "sgd": { "label": "稳定…训练", "note": "随机梯度太大时可先裁剪。" }, "lstm": { "label": "保护…训练", "note": "循环模型训练常遇到梯度爆炸。" } } } } }, { "id": "gradient-descent", "name": "Gradient Descent", "layer": "L2", "era": "1951", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "backpropagation" }, { "to": "sgd" }, { "to": "adam" }, { "to": "parameter" } ], "track": "history", "i18n": { "en": { "fullName": "Gradient Descent", "factExplain": "A method for updating parameters toward lower loss.", "humanExplain": "Gradient descent is like finding the lowest spot on a mini-golf green. You tap the ball downhill, not into next Tuesday.\n\nIt trains models by lowering error. Big steps can overshoot. Tiny steps crawl.", "humanExplainDisplay": "Gradient descent is like finding\n==the lowest spot== on a mini-golf green.\nYou tap the ball ==downhill==,\nnot into next Tuesday.\n\nIt trains models by lowering error.\nBig steps can overshoot.\nTiny steps crawl.", "relationsNarrative": "Backpropagation\nBackprop calculates the gradient, and gradient descent updates the parameters with it.\n\nSGD\nSGD is a random mini-batch version of gradient descent, and it is very common.\n\nAdam\nAdam improves gradient descent by changing the step size and update style.\n\nParameter\nGradient descent moves parameters in a better direction.", "relations": { "backpropagation": { "label": "gets direction from …", "note": "Backprop first works out which way the parameters should move." }, "sgd": { "label": "has … variant", "note": "SGD is a common version that uses less compute." }, "adam": { "label": "has … upgrade", "note": "Adam adjusts each update step on its own." }, "parameter": { "label": "updates …", "note": "Gradient descent directly changes the parameters." } } }, "zh": { "fullName": "梯度下降", "factExplain": "沿损失函数下降方向更新参数的优化方法。", "humanExplain": "像蒙眼拧水龙头试水温：先摸出偏烫偏凉，再朝更舒服那边一点点回调。\n\n常用于训练模型降误差；步子太大易过头，太小又会变慢。", "humanExplainDisplay": "像蒙眼拧水龙头试水温：\n先摸出==偏烫偏凉==，\n再朝==更舒服==那边\n一点点回调。\n\n常用于训练模型降误差；\n步子太大易过头，\n太小又会变慢。", "relationsNarrative": "Backpropagation\n反向传播负责算梯度，它负责按梯度更新参数。\n\nSGD\nSGD 是梯度下降的随机小批量版本，更常见。\n\nAdam\nAdam 在它基础上改进步长与更新方式。\n\nParameter\n梯度下降的目标，就是把参数往更优方向调。", "relations": { "backpropagation": { "label": "接收…给的方向", "note": "反向传播先算出该往哪边改。" }, "sgd": { "label": "发展出…变体", "note": "SGD 是它更省算力的常见版本。" }, "adam": { "label": "有…升级版", "note": "Adam 会自适应调节每步更新。" }, "parameter": { "label": "用来更新…", "note": "它的直接作用对象就是参数。" } } } } }, { "id": "graph-convolutional-network", "name": "Graph Convolutional Network", "layer": "L3", "era": "2017", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "cnn" }, { "to": "embedding" }, { "to": "knowledge-representation" } ], "track": "history", "i18n": { "en": { "fullName": "Graph Convolutional Network", "factExplain": "A neural network that learns from graph links and builds node representations.", "humanExplain": "A Graph Convolutional Network is like judging a house by its neighborhood. The shops, the streets, the foot traffic tell you more than the floor plan alone.\n\nIt learns from connections, not just single items. You meet it in social apps, recommendations, and molecule research.", "humanExplainDisplay": "A Graph Convolutional Network is like\njudging a house by its ==neighborhood==.\nThe shops, the streets, the foot traffic\ntell you more than ==the floor plan== alone.\n\nIt learns from connections,\nnot just single items.\nYou meet it in social apps,\nrecommendations,\nand molecule research.", "relationsNarrative": "Neural-network\nA Graph Convolutional Network is a neural network built for graph data.\n\nCNN\nIt borrows the convolution idea, but its neighbors are graph links, not grid squares.\n\nEmbedding\nIt turns nodes or whole graphs into vectors computers can use.\n\nKR\nIt helps model things and the links between them.", "relations": { "neural-network": { "label": "is a kind of …", "note": "It is still a neural network at heart." }, "cnn": { "label": "borrows ideas from …", "note": "It moves convolution from grid data onto graphs." }, "embedding": { "label": "learns … representations", "note": "It outputs vectors for nodes or whole graphs." }, "knowledge-representation": { "label": "works with … structures", "note": "It fits networks of things and their links." } } }, "zh": { "fullName": "图卷积网络", "factExplain": "一种在图结构数据上传播并学习节点表示的神经网络。", "humanExplain": "图卷积网络像看房不光看户型：邻居街区一起掂量，旁边铺子人气决定房子的真实分量。\n\n常用于社交网络、推荐和分子分析，适合处理连接关系。", "humanExplainDisplay": "图卷积网络像看房不光看户型，\n还得把邻居街区一起==掂量==：\n旁边是什么铺子、什么人流，\n决定了这房子的==真实分量==。\n\n常用于社交网络、\n推荐和分子分析，\n适合处理连接关系。", "relationsNarrative": "Neural-network\n它是专门处理图结构数据的神经网络变体。\n\nCNN\n它借鉴卷积思想，但邻域不再是规则网格。\n\nEmbedding\n它会把节点或整张图压成可计算的向量表示。\n\nKnowledge-representation\n它常用来建模实体之间的关系与连接结构。", "relations": { "neural-network": { "label": "属于…一类", "note": "它本质上仍是神经网络。" }, "cnn": { "label": "借鉴…思路", "note": "把卷积从网格数据搬到图上。" }, "embedding": { "label": "学习…表示", "note": "核心产出是节点或图的向量表示。" }, "knowledge-representation": { "label": "处理…结构", "note": "适合建模实体及其关系网络。" } } } } }, { "id": "graph-neural-network", "name": "GNN", "layer": "L3", "era": "2005", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "graph-convolutional-network" }, { "to": "knowledge-graph" }, { "to": "embedding" } ], "track": "history", "i18n": { "en": { "fullName": "Graph Neural Network", "factExplain": "A neural network built to learn from dots and links.", "humanExplain": "A GNN is the lunchroom rumor expert. It studies who sits together, then spots the fry thief faster.\n\nIt works on data made of things and their links. You meet it in friend maps, recommendations, and molecule studies.", "humanExplainDisplay": "A GNN is the ==lunchroom rumor expert==.\nIt studies ==who sits together==,\nthen spots the fry thief faster.\n\nIt works on data made of things\nand their links.\nYou meet it in friend maps,\nrecommendations,\nand molecule studies.", "relationsNarrative": "Neural-network\nA GNN is a Neural-network built for relationship maps.\n\nGraph Convolutional Network\nGraph Convolutional Network is a classic type of GNN.\n\nKnowledge Graph\nA GNN can model a Knowledge Graph and reason over its links.\n\nEmbedding\nA GNN turns nodes and links into Embeddings for prediction.", "relations": { "neural-network": { "label": "is a kind of …", "note": "It is a neural network built for graph data." }, "graph-convolutional-network": { "label": "includes classic …", "note": "Graph Convolutional Network is a common type of GNN." }, "knowledge-graph": { "label": "models … structures", "note": "A Knowledge Graph is often a target for GNN modeling." }, "embedding": { "label": "turns nodes into …", "note": "It learns vector forms for nodes and links." } } }, "zh": { "fullName": "Graph Neural Network｜图神经网络", "factExplain": "一种专门处理图结构数据的神经网络。", "humanExplain": "GNN 看人不只看简历，还顺手打听同学、同事、前老板，关系网越清越有数。\n\n常用于社交网络、推荐和分子结构分析，谁和谁相连，本身就是重要信息。", "humanExplainDisplay": "GNN 看人不只看简历，\n还顺手打听==同学、同事、前老板==，\n==关系网越清越有数==。\n\n常用于社交网络、\n推荐和分子结构分析，\n谁和谁相连，本身就是重要信息。", "relationsNarrative": "Neural-network\n它是神经网络家族里，专门处理关系网的一支。\n\nGraph Convolutional Network\n图卷积网络是图神经网络里最经典的变体之一。\n\nKnowledge Graph\n知识图谱这类关系网络，常用它来做建模和推理。\n\nEmbedding\n它会把节点关系压成向量，便于分类和预测。", "relations": { "neural-network": { "label": "属于…一类", "note": "它是面向图数据的神经网络分支。" }, "graph-convolutional-network": { "label": "包含经典变体…", "note": "图卷积网络是它最常见的一支。" }, "knowledge-graph": { "label": "可建模…结构", "note": "知识图谱常是它的重要应用对象。" }, "embedding": { "label": "把节点变成…", "note": "它会学习节点和边的向量表示。" } } } } }, { "id": "graph-search", "name": "Graph Search", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "heuristic-search" }, { "to": "dynamic-programming" }, { "to": "agent" }, { "to": "world-model" } ], "track": "history", "i18n": { "en": { "fullName": "Graph Search", "factExplain": "A way to find a path or target in a network of connected points.", "humanExplain": "Graph Search is like finding your car in a huge parking garage. You check one lane at a time, not all of Level 3 in a panic.\n\nIt helps plan routes and explore possible states. Many Agents use it to choose what to do next.", "humanExplainDisplay": "Graph Search is like finding your car\nin a ==huge parking garage==.\nYou check ==one lane at a time==,\nnot all of Level 3 in a panic.\n\nIt helps plan routes\nand explore possible states.\nMany Agents use it\nto choose what to do next.", "relationsNarrative": "Heuristic Search\nHeuristic Search adds helpful clues to Graph Search.\n\nDP\nDP can reuse repeated pieces inside a graph.\n\nAgent\nMany Agents turn a task into states and search for a path.\n\nWorld model\nA World model gives Graph Search states and moves to explore.", "relations": { "heuristic-search": { "label": "can be sped up by …", "note": "Heuristics help Graph Search move toward the goal faster." }, "dynamic-programming": { "label": "shares ideas with …", "note": "DP often reuses answers from repeated parts of a graph." }, "agent": { "label": "plans for …", "note": "Many Agents search a state graph to find an action path." }, "world-model": { "label": "searches inside …", "note": "A World model gives it states and moves to search." } } }, "zh": { "fullName": "图搜索", "factExplain": "在图结构中按规则寻找路径或目标节点的方法。", "humanExplain": "像在大型景区找出口：不是满园乱窜，而是沿着岔路一口口试，排查哪条道真能通出去。\n\n常用于路径规划和状态探索，是很多智能体做决策的基础。", "humanExplainDisplay": "像在大型景区找出口：\n不是==满园乱窜==，\n而是沿着岔路一口口试，\n排查哪条道==真能通出去==。\n\n常用于路径规划和状态探索，\n是很多智能体做决策的基础。", "relationsNarrative": "Heuristic-search\n启发式搜索是在图搜索上加入经验指路。\n\nDynamic-programming\n动态规划常利用图中重叠子问题做高效求解。\n\nAgent\n很多智能体会把任务拆成状态图再搜索。\n\nWorld-model\n世界模型提供可搜索的状态与转移结构。", "relations": { "heuristic-search": { "label": "可被…加速", "note": "启发式能帮它更快逼近目标。" }, "dynamic-programming": { "label": "与…思路相邻", "note": "两者都在系统利用子问题结构。" }, "agent": { "label": "为…提供规划", "note": "很多智能体靠它寻找行动路径。" }, "world-model": { "label": "在…上展开搜索", "note": "有内部世界模型时更好做规划。" } } } } }, { "id": "grok-4-5", "name": "Grok 4.5", "layer": "L3", "era": "2026", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "frontier-model" }, { "to": "reasoning-model" } ], "track": "today", "i18n": { "en": { "fullName": "xAI Grok 4.5 Large Language Model", "factExplain": "An xAI version of the Grok large language model family.", "humanExplain": "Grok 4.5 is the science kid at lunch. It brings a calculator and a fresh meme. It may crack a joke too.\n\nYou meet it in chat, code, and search. Test it on your own task first.", "humanExplainDisplay": "Grok 4.5 is the ==science kid at lunch==.\nIt brings a calculator\nand a ==fresh meme==.\nIt may crack a joke too.\n\nYou meet it in chat, code, and search.\nTest it on your own task first.", "relationsNarrative": "LLM\nGrok 4.5 is one product version of an LLM.\n\nFrontier model\nGrok 4.5 competes in the frontier model race.\n\nReasoning-model\nIts hard-task results depend on reasoning skill.", "relations": { "llm": { "label": "is a version of …", "note": "Language understanding and writing are its core." }, "frontier-model": { "label": "competes in … race", "note": "It sits in the race for top general AI models." }, "reasoning-model": { "label": "builds stronger … skills", "note": "Reasoning shapes how well it handles hard tasks." } } }, "zh": { "fullName": "xAI Grok 4.5 大语言模型", "factExplain": "xAI 发布的 Grok 系列大语言模型版本。", "humanExplain": "Grok 4.5像班里嘴快的理科同桌：会刷题、追热搜，偶尔还贫你两句。\n\n用于聊天、代码和搜索，表现要按任务实测。", "humanExplainDisplay": "Grok 4.5像班里\n==嘴快的理科同桌==：\n会刷题、追热搜，\n偶尔还==贫你两句==。\n\n用于聊天、代码和搜索，\n表现要按任务实测。", "relationsNarrative": "LLM\nGrok 4.5 是 LLM 的一个具体产品版本。\n\nFrontier Model\n它参与前沿模型的能力竞赛。\n\nReasoning Model\n它的复杂任务表现依赖推理能力。", "relations": { "llm": { "label": "属于…", "note": "它以语言理解和生成为核心。" }, "frontier-model": { "label": "参与…竞赛", "note": "它定位在高端通用模型赛道。" }, "reasoning-model": { "label": "强化…能力", "note": "推理能力决定复杂任务表现。" } } } } }, { "id": "groq", "name": "Groq", "layer": "L5", "sublayer": "compute", "era": "2016", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "ai-chip" }, { "to": "llm-inference-engine" }, { "to": "tokens-per-second" }, { "to": "api" } ], "track": "history", "i18n": { "en": { "fullName": "Groq", "factExplain": "A company that offers fast AI chips and cloud service for LLM answers.", "humanExplain": "Groq is like giving an AI chatbot a race-car pit crew. Same brain. Much faster replies.\n\nYou meet it behind fast chat and voice apps. It aims for low delay and many answers at once.", "humanExplainDisplay": "Groq is like giving an AI chatbot\na ==race-car pit crew==.\nSame brain.\n==Much faster replies==.\n\nYou meet it behind fast chat and voice apps.\nIt aims for low delay\nand many answers at once.", "relationsNarrative": "AI chip\nGroq uses special AI chips to speed up model answers.\n\nInference engine\nGroq focuses on running models, not training them.\n\nTPS\nTPS often shows Groq's speed advantage.\n\nAPI\nGroq mainly delivers fast inference through an API.", "relations": { "ai-chip": { "label": "speeds inference with …", "note": "Special AI chips are Groq's main selling point." }, "llm-inference-engine": { "label": "serves …", "note": "Groq focuses on running models, not training them." }, "tokens-per-second": { "label": "chases higher …", "note": "TPS is a common way to measure output speed." }, "api": { "label": "delivers through …", "note": "Developers usually call Groq through an API." } } }, "zh": { "fullName": "Groq（AI 推理芯片与云服务公司）", "factExplain": "一家提供高速 LLM 推理芯片与云服务的公司。", "humanExplain": "Groq 像给 AI 装上高铁轮子：脑子没换，回答却嗖嗖出站。\n\n用于高速聊天、语音和批量推理，主打低延迟高吞吐。", "humanExplainDisplay": "Groq 像给 AI\n==装上高铁轮子==：\n脑子没换，\n回答却==嗖嗖出站==。\n\n用于高速聊天、语音\n和批量推理，\n主打低延迟高吞吐。", "relationsNarrative": "AI chip\nGroq 的核心是用专用芯片加速模型推理。\n\nInference engine\nGroq 面向推理服务，重点不是训练模型。\n\nTPS\nTPS 常用来衡量 Groq 的输出速度优势。\n\nAPI\nGroq 主要通过 API 把高速推理交给开发者。", "relations": { "ai-chip": { "label": "用…加速推理", "note": "专用芯片是它的核心卖点。" }, "llm-inference-engine": { "label": "服务…", "note": "它面向模型推理而非训练。" }, "tokens-per-second": { "label": "追求更高…", "note": "输出速度常用 TPS 衡量。" }, "api": { "label": "通过…交付", "note": "开发者多用接口调用服务。" } } } } }, { "id": "gru", "name": "GRU", "layer": "L3", "era": "2014", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "recurrent-neural-network" }, { "to": "lstm" }, { "to": "seq2seq" }, { "to": "backpropagation-through-time" } ], "track": "history", "i18n": { "en": { "fullName": "Gated Recurrent Unit", "factExplain": "A gated RNN that remembers useful sequence details and forgets the rest.", "humanExplain": "GRU is like a student with a tiny backpack. It keeps the test hints and dumps the cafeteria gossip.\n\nYou meet it in speech, translation, and time data. It remembers enough, but stays lighter than LSTM.", "humanExplainDisplay": "GRU is like a student with a ==tiny backpack==.\nIt keeps the ==test hints==\nand dumps the cafeteria gossip.\n\nYou meet it in speech, translation, and time data.\nIt remembers enough,\nbut stays lighter than LSTM.", "relationsNarrative": "RNN\nGRU is a gated RNN that helps reduce lost long-term memory.\n\nLSTM\nGRU is simpler than LSTM, so it often trades some power for speed.\n\nSeq2Seq\nEarly Seq2Seq translation often used GRUs to encode and decode sequences.\n\nBPTT\nGRU is usually trained with BPTT after the sequence is unrolled over time.", "relations": { "recurrent-neural-network": { "label": "improves …", "note": "GRU is a gated upgrade of an RNN." }, "lstm": { "label": "simplifies …", "note": "GRU uses fewer gates for lighter memory." }, "seq2seq": { "label": "supports …", "note": "Early translation models often used GRUs to encode sequences." }, "backpropagation-through-time": { "label": "trains with …", "note": "BPTT sends sequence errors backward through time." } } }, "zh": { "fullName": "Gated Recurrent Unit，门控循环单元", "factExplain": "一种用门控机制处理序列记忆的循环神经网络。", "humanExplain": "GRU像会抓重点的课堂学霸：板书抄下，闲聊随风飘走。\n\n用于语音、翻译、时间序列，记忆能力够，而且更轻。", "humanExplainDisplay": "GRU像会抓重点的课堂学霸：\n==板书抄下==，\n==闲聊随风飘走==。\n\n用于语音、翻译、时间序列，\n记忆能力够，\n而且更轻。", "relationsNarrative": "RNN\nGRU 是 RNN 的门控变体，用来缓解长期记忆丢失。\n\nLSTM\nGRU 比 LSTM 结构更简洁，常在效果和速度间折中。\n\nSeq2Seq\n早期 Seq2Seq 翻译常用 GRU 编码和解码序列。\n\nBPTT\nGRU 通常通过 BPTT 在时间展开后训练。", "relations": { "recurrent-neural-network": { "label": "改造…", "note": "它是 RNN 的门控改良版。" }, "lstm": { "label": "简化…", "note": "它用更少门控换取轻量记忆。" }, "seq2seq": { "label": "支撑…", "note": "早期翻译模型常用它编码序列。" }, "backpropagation-through-time": { "label": "依靠…训练", "note": "序列误差通过时间反向传播。" } } } } }, { "id": "hallucination", "name": "Hallucination", "layer": "L4", "era": "2023", "publishedAt": "2026-05-23T08:10:00Z", "relations": [ { "to": "llm" }, { "to": "prompt" }, { "to": "rag" }, { "to": "alignment" } ], "i18n": { "en": { "fullName": "AI Hallucination", "factExplain": "An AI answer that sounds right but is actually false or unproven.", "humanExplain": "Hallucination is like a friend giving directions with full confidence. Then you find the pizza place is a pond.\n\nIt can fill blanks when writing. It can trap you during research. Check important facts before you trust it.", "humanExplainDisplay": "Hallucination is like a friend giving directions\nwith ==full confidence==.\nThen you find the ==pizza place is a pond==.\n\nIt can fill blanks when writing.\nIt can trap you during research.\nCheck important facts before you trust it.", "relationsNarrative": "LLM\nAn LLM predicts likely words, so it can hallucinate.\n\nPrompt\nA clear Prompt leaves less empty space for hallucination.\n\nRAG\nRAG gives the AI sources it can check before it answers.\n\nAlignment\nAlignment trains the AI to avoid shaky answers.", "relations": { "llm": { "label": "is common in …", "note": "LLMs can invent facts because they write likely-sounding text." }, "prompt": { "label": "shrinks with clear …", "note": "A clear prompt gives the AI less room to guess." }, "rag": { "label": "is reduced by …", "note": "RAG gives the AI sources to check before answering." }, "alignment": { "label": "is trained down by …", "note": "Alignment teaches the AI to avoid unsafe or shaky answers." } } }, "zh": { "fullName": "幻觉", "factExplain": "AI 生成看似合理但实际错误或无法验证的信息。", "humanExplain": "幻觉像一个死不认输的同事，不知道答案也能编得像开过会。\n\n写作时它能补空，查资料时它能挖坑；越关键的信息，越不能只听它一张嘴。", "humanExplainDisplay": "幻觉像一个\n==死不认输的同事==。\n不知道答案，也能编得像刚开完会。\n\n写作时它能补空，\n查资料时它能挖坑。\n越关键的信息，越别只听它一张嘴。", "relationsNarrative": "LLM\nLLM 以概率方式生成内容，因此可能出现 Hallucination。\n\nPrompt\nPrompt 越明确，Hallucination 可发挥的模糊空间越小。\n\nRAG\nRAG 引入外部资料，为模型回答提供可核查依据。\n\nAlignment\nAlignment 通过约束模型行为，减少不可靠输出。", "relations": { "llm": { "label": "是…的常见问题" }, "prompt": { "label": "可被清晰…缓解" }, "rag": { "label": "靠…降低" }, "alignment": { "label": "靠…训练抑制" } } } } }, { "id": "heuristic-search", "name": "Heuristic Search", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "graph-search" }, { "to": "agent" }, { "to": "reasoning-model" }, { "to": "dynamic-programming" } ], "track": "history", "i18n": { "en": { "fullName": "Heuristic Search", "factExplain": "A search method uses smart guesses to find good answers faster.", "humanExplain": "Heuristic Search is like hunting for your car at the mall. You try the row near the big lamp, not every row like a confused Roomba.\n\nIt uses quick guesses to pick the next path in planning and puzzle solving. It saves time, but it may miss the best answer.", "humanExplainDisplay": "Heuristic Search is like ==hunting for your car== at the mall.\nYou try the row near the ==big lamp==,\nnot every row like a confused Roomba.\n\nIt uses quick guesses to pick the next path\nin planning and puzzle solving.\nIt saves time,\nbut it may miss the best answer.", "relationsNarrative": "Graph Search\nHeuristic Search uses quick scoring rules to choose which path to try first.\n\nAgent\nAn Agent can use Heuristic Search to plan steps and avoid dead ends.\n\nReasoning-model\nA Reasoning-model can use Heuristic Search to shrink the search space.\n\nDP\nHeuristic Search guides the search, while DP reuses solved smaller problems.", "relations": { "graph-search": { "label": "adds guesses to …", "note": "It adds quick scoring rules to search." }, "agent": { "label": "helps … find steps", "note": "An Agent can use it to plan a path of actions." }, "reasoning-model": { "label": "helps … plan", "note": "It can shrink the search space during reasoning." }, "dynamic-programming": { "label": "is often compared with …", "note": "One guides search; the other reuses solved parts." } } }, "zh": { "fullName": "启发式搜索", "factExplain": "用估价规则引导搜索，更快找到较优解的方法。", "humanExplain": "像老中医看人先搭脉摸门道，不会把所有方子都试一遍，而是先朝最像病根的方向下手。\n\n常用于规划和求解，能省大量时间，但不保证次次最优。", "humanExplainDisplay": "像老中医看人先==搭脉摸门道==，\n不会把所有方子都试一遍，\n而是先朝最像==病根==的方向下手。\n\n常用于规划和求解，\n能省大量时间，\n但不保证次次最优。", "relationsNarrative": "Graph-search\n它是在图搜索基础上，用估价规则优先扩展方向。\n\nAgent\n智能体可借它规划任务步骤，少走无效路径。\n\nReasoning-model\n推理模型可借它缩小搜索范围，提高求解效率。\n\nDynamic-programming\n它常和动态规划对照：一个重引导搜索，一个重子问题复用。", "relations": { "graph-search": { "label": "属于…思路", "note": "它是在搜索里加估价规则。" }, "agent": { "label": "帮助…找步骤", "note": "智能体可用它规划行动路径。" }, "reasoning-model": { "label": "可为…做规划", "note": "推理时可用来缩小搜索范围。" }, "dynamic-programming": { "label": "常与…对照", "note": "一个偏搜索，一个偏系统拆解。" } } } } }, { "id": "hidden-markov-model", "name": "HMM", "layer": "L2", "era": "1966", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "markov-decision-process" }, { "to": "bayesian-network" }, { "to": "speech-to-text" }, { "to": "lstm" } ], "track": "history", "i18n": { "en": { "fullName": "Hidden Markov Model", "factExplain": "A probability model that guesses hidden states from a visible sequence.", "humanExplain": "HMM is like guessing the weather from your brother’s sneakers each morning. Muddy shoes whisper, “Rain,” like tiny, dirty tattletales.\n\nIt follows clues in order and guesses the hidden state behind each one. You meet it in speech tools and sequence tagging.", "humanExplainDisplay": "HMM is like guessing the weather\nfrom your brother’s ==sneakers== each morning.\n==Muddy shoes== whisper, “Rain,”\nlike tiny, dirty tattletales.\n\nIt follows clues in order\nand guesses the hidden state behind each one.\nYou meet it in speech tools\nand sequence tagging.", "relationsNarrative": "MDP\nBoth use the Markov idea, but HMM does not choose actions.\n\nBayesian Network\nHMM can be seen as a Bayesian Network repeated over time.\n\nSTT\nHMMs were long used in STT to match sounds to words.\n\nLSTM\nLSTMs later replaced HMMs in many sequence tasks.", "relations": { "markov-decision-process": { "label": "shares the Markov idea with …", "note": "Both assume the current state carries the useful past." }, "bayesian-network": { "label": "is a time-based cousin of …", "note": "It can be seen as a Bayesian Network repeated over time." }, "speech-to-text": { "label": "was often used in …", "note": "Early STT used HMMs to decode sound into words." }, "lstm": { "label": "was partly replaced by …", "note": "Later deep sequence models took over many of its jobs." } } }, "zh": { "fullName": "Hidden Markov Model（隐马尔可夫模型）", "factExplain": "一种用隐藏状态生成可观测序列的概率模型。", "humanExplain": "HMM像听隔壁屋打麻将：你只听见碰杠胡，屋里谁手气正旺，还得顺着前后动静慢慢猜。\n\n它根据已看到的信号，推断背后状态，常用于语音和序列标注。", "humanExplainDisplay": "HMM像听隔壁屋打麻将：\n你只听见==碰杠胡==，\n屋里谁手气正旺，\n还得==顺着前后动静慢慢猜==。\n\n它根据已看到的信号，\n推断背后状态；\n常用于语音和序列标注。", "relationsNarrative": "Markov Decision Process\n两者都建立在马尔可夫假设上，但它不负责做决策。\n\nBayesian Network\n它可看作一种按时间展开的动态概率图模型。\n\nSpeech-to-text\n它曾长期用于语音识别里的声学建模和解码。\n\nLSTM\n在很多序列任务里，LSTM 后来逐步替代了它。", "relations": { "markov-decision-process": { "label": "共享马尔可夫假设", "note": "两者都假设当前状态概括过去。" }, "bayesian-network": { "label": "属于…近亲", "note": "它可看作按时间展开的概率图模型。" }, "speech-to-text": { "label": "常被用于…", "note": "早期语音识别常靠它做序列解码。" }, "lstm": { "label": "被…部分取代", "note": "深度序列模型后来覆盖了许多场景。" } } } } }, { "id": "hidream-o1-image-1-5", "name": "HiDream-O1-Image-1.5", "layer": "L3", "era": "2026", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "multimodal" }, { "to": "prompt" }, { "to": "seedance" } ], "track": "history", "i18n": { "en": { "fullName": "HiDream-O1-Image-1.5", "factExplain": "An AI image model that makes pictures from text prompts.", "humanExplain": "HiDream-O1 is like the doodle kid in art class. Say “robot panda selling boba,” and it starts sketching.\n\nPeople use it for posters and illustration work. It makes concept pictures fast, but your prompt still steers the result.", "humanExplainDisplay": "HiDream-O1 is like the ==doodle kid==\nin art class.\nSay ==“robot panda selling boba,”==\nand it starts sketching.\n\nPeople use it for posters\nand illustration work.\nIt makes concept pictures fast,\nbut your prompt still steers the result.", "relationsNarrative": "Diffusion\nHiDream-O1 usually uses diffusion to make images.\n\nMultimodal AI\nHiDream-O1 reads words and turns them into pictures.\n\nPrompt\nA prompt is its direct input and steers the picture.\n\nSeedance\nHiDream-O1 makes images, while Seedance makes video.", "relations": { "diffusion": { "label": "follows …", "note": "Most text-to-image models use this kind of method." }, "multimodal": { "label": "is a kind of …", "note": "It works across words and pictures." }, "prompt": { "label": "draws from …", "note": "Your wording often decides the picture's direction." }, "seedance": { "label": "shares a family with …", "note": "HiDream-O1 makes images, while Seedance makes video." } } }, "zh": { "fullName": "HiDream-O1-Image-1.5 图像生成模型", "factExplain": "一个用于根据文本生成图片的 AI 图像模型。", "humanExplain": "像同桌偷摸画课本涂鸦，你刚念一句“机甲熊猫卖奶茶”，它唰唰几笔，离谱画面还真给你整出来。\n\n常用于海报、插画和概念图生成，出图很快，但画面效果仍很依赖提示词。", "humanExplainDisplay": "像同桌偷摸画课本涂鸦，\n你刚念一句“机甲熊猫卖奶茶”，\n它唰唰几笔，离谱画面还真给你\n==整出来==。\n\n常用于海报、\n插画和概念图生成，\n出图很快，但画面效果仍很依赖提示词。", "relationsNarrative": "Diffusion\n它通常属于扩散式图像生成这一路线。\n\nMultimodal\n它把文字理解后转成图像，属于多模态生成。\n\nPrompt\n提示词是它出图的直接输入，决定画面方向。\n\nSeedance\n两者同属生成模型家族，但一个主打图片，一个主打视频。", "relations": { "diffusion": { "label": "属于…路线", "note": "大多数文生图模型都建立在这类方法上。" }, "multimodal": { "label": "是…的一种", "note": "它处理文字到图像的跨模态生成。" }, "prompt": { "label": "靠…出图", "note": "用户怎么描述，往往决定画面方向。" }, "seedance": { "label": "同属生成家族", "note": "一个偏图像，一个偏视频生成。" } } } } }, { "id": "hierarchical-reinforcement-learning", "name": "HRL", "layer": "L2", "era": "1990s", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "options-framework" }, { "to": "markov-decision-process" }, { "to": "long-horizon-task" } ], "track": "history", "i18n": { "en": { "fullName": "Hierarchical Reinforcement Learning", "factExplain": "HRL breaks long reinforcement learning jobs into higher goals and lower skills.", "humanExplain": "HRL is like cleaning your room with a tiny boss in your head. The boss says “make it livable,” then sends smaller helpers for socks and scary cups.\n\nRobots and game AIs use it for long jobs. It cuts trial and error by reusing small learned skills.", "humanExplainDisplay": "HRL is like cleaning your room\nwith a ==tiny boss== in your head.\nThe boss says “make it livable,”\nthen sends ==smaller helpers==\nfor socks and scary cups.\n\nRobots and game AIs use it for long jobs.\nIt cuts trial and error\nby reusing small learned skills.", "relationsNarrative": "RL\nHRL is a layered way to use RL on long tasks.\n\nOptions Framework\nThe Options Framework gives HRL reusable sub-policies.\n\nMDP\nHRL usually uses an MDP for states, actions, and rewards.\n\nLong-horizon\nHRL breaks long tasks into levels, so they are easier to explore.", "relations": { "reinforcement-learning": { "label": "adds layers to …", "note": "HRL splits RL into high-level and low-level policies." }, "options-framework": { "label": "organizes skills with …", "note": "Options are reusable sub-policies for HRL." }, "markov-decision-process": { "label": "defines tasks with …", "note": "The MDP base still supplies states, actions, and rewards." }, "long-horizon-task": { "label": "breaks down …", "note": "Layers make the search path shorter for long tasks." } } }, "zh": { "fullName": "分层强化学习", "factExplain": "把长期任务拆成层级策略来学习的强化学习方法。", "humanExplain": "HRL像拍电影分工：导演管全片，场记拆镜头，演员一场场演。\n\n用于机器人和游戏长任务，减少探索成本，复用子策略。", "humanExplainDisplay": "HRL像拍电影分工：\n导演管全片，\n==场记拆镜头==，\n演员==一场场演==。\n\n用于机器人和游戏长任务，\n减少探索成本，\n复用子策略。", "relationsNarrative": "RL\nHRL 是强化学习处理长任务的一种分层思路。\n\nOptions Framework\nOptions Framework 为 HRL 定义可复用子策略。\n\nMDP\nHRL 通常仍在 MDP 框架下描述状态与奖励。\n\nLong-horizon\nHRL 用层级拆解缓解长程任务难探索。", "relations": { "reinforcement-learning": { "label": "分层扩展…", "note": "把强化学习拆成上下级策略。" }, "options-framework": { "label": "用…组织子策略", "note": "子策略是层级学习的常见积木。" }, "markov-decision-process": { "label": "基于…定义任务", "note": "状态、动作、奖励仍是底座。" }, "long-horizon-task": { "label": "拆解…", "note": "层级拆解能缩短探索路径。" } } } } }, { "id": "hopfield-network", "name": "Hopfield Network", "layer": "L3", "era": "1982", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "recurrent-neural-network" }, { "to": "boltzmann-machine" }, { "to": "connectionism" } ], "track": "history", "i18n": { "en": { "fullName": "Hopfield Network", "factExplain": "A recurrent neural network used for associative memory.", "humanExplain": "A Hopfield Network is like a friend hearing your awful humming. Three wobbly notes later, they still guess the song.\n\nIt stores patterns and pulls messy inputs back to a close match. You meet it in memory ideas and image cleanup.", "humanExplainDisplay": "A Hopfield Network is like a friend\nhearing your ==awful humming==.\nThree ==wobbly notes== later,\nthey still guess the song.\n\nIt stores patterns\nand pulls messy inputs back to a close match.\nYou meet it in memory ideas\nand image cleanup.", "relationsNarrative": "RNN\nIt is an early RNN, and it repeats updates to find a memory.\n\nEnergy-Based Model\nIt uses energy, and remembered patterns sit in low spots.\n\nBoltzmann Machine\nThe Boltzmann Machine kept its energy view and added random sampling.\n\nConnectionism\nIt became a key network in the comeback of Connectionism.", "relations": { "recurrent-neural-network": { "label": "is a kind of …", "note": "It is a classic early RNN." }, "boltzmann-machine": { "label": "inspired …", "note": "Boltzmann Machines kept the energy idea and added chance." }, "connectionism": { "label": "helped revive …", "note": "It helped make connectionist networks exciting again." } } }, "zh": { "fullName": "霍普菲尔德网络", "factExplain": "一种用于联想记忆的循环神经网络。", "humanExplain": "Hopfield网络像KTV猜歌：前奏跑调几拍，也能把旋律拉回原曲。\n\n用于联想记忆和图像纠错，也启发能量模型。", "humanExplainDisplay": "Hopfield网络像KTV猜歌：\n前奏==跑调几拍==，\n也能把旋律\n==拉回原曲==。\n\n用于联想记忆\n和图像纠错，\n也启发能量模型。", "relationsNarrative": "RNN\n它是早期循环网络，用状态反复更新记忆。\n\nEnergy-based Model\n它用能量函数把记忆变成可收敛的低谷。\n\nBoltzmann Machine\n玻尔兹曼机继承能量视角，并加入随机采样。\n\nConnectionism\n它是连接主义复兴中的标志性神经网络。", "relations": { "recurrent-neural-network": { "label": "属于…", "note": "它是早期循环神经网络代表。" }, "boltzmann-machine": { "label": "启发…", "note": "玻尔兹曼机延续能量网络思路。" }, "connectionism": { "label": "推动…", "note": "它让连接主义重新受到关注。" } } } } }, { "id": "hough-transform", "name": "Hough Transform", "layer": "L4", "era": "1962", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "ransac" }, { "to": "feature-engineering" }, { "to": "object-detection" } ], "track": "history", "i18n": { "en": { "fullName": "Hough Transform", "factExplain": "A method that lets image points vote in parameter space to find shapes.", "humanExplain": "Hough Transform is a school election for hidden lines. Every edge dot votes, and the line with the most votes gets the crown.\n\nIt finds lane lines and round parts. It is old-school, but it can still spot position marks in noisy images.", "humanExplainDisplay": "Hough Transform is a ==school election==\nfor hidden lines.\nEvery edge dot votes,\nand the line with the ==most votes==\ngets the crown.\n\nIt finds lane lines and round parts.\nIt is old-school,\nbut it can still spot position marks\nin noisy images.", "relationsNarrative": "Computer Vision\nHough Transform is a classic way to find lines and circles in images.\n\nRANSAC\nBoth find shapes among noisy points, but they use different tricks.\n\nFeature-engineering\nIt turns edge points into handmade geometry features.\n\nObject Detection\nEarly object detection used votes to find shapes and parts.", "relations": { "computer-vision": { "label": "finds shapes for …", "note": "Lines and circles are classic vision clues." }, "ransac": { "label": "fits noisy shapes like …", "note": "RANSAC samples points; Hough Transform counts votes." }, "feature-engineering": { "label": "gives clues for …", "note": "Handmade geometry features used to work very well." }, "object-detection": { "label": "helps … locate objects", "note": "Shape votes can help find an object's place." } } }, "zh": { "fullName": "霍夫变换", "factExplain": "把图像点投票到参数空间以检测几何形状。", "humanExplain": "霍夫变换像菜市场选摊王：边缘点挨个投票，票最多的那条线当选。\n\n常找车道线、圆形零件和定位标记，传统但抗噪。", "humanExplainDisplay": "霍夫变换像菜市场\n==选摊王==：\n边缘点==挨个投票==，\n票最多的那条线==当选==。\n\n常找车道线、圆形零件\n和定位标记，\n传统但抗噪。", "relationsNarrative": "Computer Vision\n它是经典视觉中检测直线、圆的基础算法。\n\nRANSAC\n两者都在噪声点里找几何模型，思路不同。\n\nFeature-engineering\n它把边缘点变成可用的手工几何特征。\n\nObject Detection\n早期目标检测常用投票定位形状和部件。", "relations": { "computer-vision": { "label": "为…检测形状", "note": "直线圆形是经典视觉线索。" }, "ransac": { "label": "和…都抗噪拟合", "note": "RANSAC靠抽样，它靠投票。" }, "feature-engineering": { "label": "提供…线索", "note": "手工几何特征曾很管用。" }, "object-detection": { "label": "辅助…定位", "note": "形状投票能辅助定位目标。" } } } } }, { "id": "hugging-face-transformers", "name": "Hugging Face Transformers", "layer": "L5", "sublayer": "product", "era": "2019", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "hugging-face" }, { "to": "transformer" }, { "to": "pytorch" }, { "to": "fine-tuning" } ], "track": "history", "i18n": { "en": { "fullName": "Hugging Face Transformers", "factExplain": "An open-source library for downloading, running, and fine-tuning pretrained AI models.", "humanExplain": "It is the school cafeteria for AI models. The popular ones are on trays. You grab one and skip cooking in a lab coat.\n\nDevelopers use it to load, run, and fine-tune pretrained models. Those models can handle text, images, or audio.", "humanExplainDisplay": "It is the ==school cafeteria== for AI models.\nThe popular ones are on trays.\nYou grab one\nand ==skip cooking== in a lab coat.\n\nDevelopers use it to load, run,\nand fine-tune pretrained models.\nThose models can handle text,\nimages, or audio.", "relationsNarrative": "Hugging Face\nHugging Face maintains this library and provides the model hub around it.\n\nTransformer\nIt wraps many Transformer models behind one common interface.\n\nPyTorch\nPyTorch often runs under it for training and inference.\n\nFine-tuning\nIt makes fine-tuning feel more like changing settings than rewriting code.", "relations": { "hugging-face": { "label": "maintained by …", "note": "Hugging Face maintains this library and hosts many models." }, "transformer": { "label": "wraps … models", "note": "Many Transformer models can be loaded with the same style of code." }, "pytorch": { "label": "runs on …", "note": "PyTorch is a common backend for training and inference." }, "fine-tuning": { "label": "supports …", "note": "It makes fine-tuning pretrained models need much less code." } } }, "zh": { "fullName": "Hugging Face Transformers 模型库", "factExplain": "用于下载、调用和微调预训练模型的开源库。", "humanExplain": "它像AI模型共享单车：热门车排好，扫码开锁就能骑走，不用自己焊车架。\n\n开发者用它加载、调用、微调多模态预训练模型。", "humanExplainDisplay": "它像\n==AI模型共享单车==：\n热门车排好，\n==扫码开锁就能骑走==。\n\n开发者用它加载、调用、\n微调\n多模态预训练模型。", "relationsNarrative": "Hugging Face\nHugging Face 维护它，并提供配套模型仓库。\n\nTransformer\n它把各种 Transformer 模型封装成统一接口。\n\nPyTorch\nPyTorch 常作为它训练和推理的底层框架。\n\nFine-tuning\n它让微调预训练模型更像改配置而非重写代码。", "relations": { "hugging-face": { "label": "由…维护", "note": "Hugging Face 负责维护这个库。" }, "transformer": { "label": "封装…模型", "note": "大量 Transformer 模型可直接加载。" }, "pytorch": { "label": "运行在…上", "note": "PyTorch 是它常用的训练后端。" }, "fine-tuning": { "label": "支持…", "note": "微调预训练模型时少写很多代码。" } } } } }, { "id": "hugging-face", "name": "Hugging Face", "layer": "L5", "sublayer": "product", "era": "2016", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "model-leaderboard" }, { "to": "api" }, { "to": "pytorch" } ], "track": "history", "i18n": { "en": { "fullName": "Hugging Face", "factExplain": "A site for sharing, downloading, and running AI models and datasets.", "humanExplain": "Hugging Face is like a huge library for AI models. The shelves have models and datasets, not dusty math books.\n\nPeople use it to find a model. They can download its weights or try it online. Many open-source models launch there first.", "humanExplainDisplay": "Hugging Face is like a ==huge library== for AI models.\nThe shelves have ==models and datasets==,\nnot dusty math books.\n\nPeople use it to find a model.\nThey can download its weights or try it online.\nMany open-source models launch there first.", "relationsNarrative": "Open-source-model\nPeople often release and download open-source models on Hugging Face.\n\nLeaderboard\nMany model leaderboards and test pages live on Hugging Face.\n\nAPI\nHugging Face lets you call some models online through an API.\n\nPyTorch\nMany models on Hugging Face are built for the PyTorch world.", "relations": { "open-source-model": { "label": "distributes …", "note": "Many open-source models are released there first." }, "model-leaderboard": { "label": "hosts …", "note": "It often shows model tests and rankings." }, "api": { "label": "offers … access", "note": "You can call hosted models online, not only download them." }, "pytorch": { "label": "connects to …", "note": "Many shared models and code use PyTorch." } } }, "zh": { "fullName": "AI 模型与数据集开源社区平台", "factExplain": "一个用于分享、下载和部署 AI 模型的平台。", "humanExplain": "逛 Hugging Face，像进大学图书馆借参考书：模型、数据集都在架上，拿到就能翻来试。\n\n常用来找模型、下权重、试效果，也是开源模型分发入口。", "humanExplainDisplay": "逛 Hugging Face，\n像进大学图书馆\n借==参考书==：\n拿到就能==翻来试==。\n\n常用来找模型、\n下权重、试效果；\n也是开源模型\n分发入口。", "relationsNarrative": "Open-source-model\n它是开源模型最常见的发布、下载和分发阵地之一。\n\nLeaderboard\n很多模型榜单和评测页面，都会放在它的平台里展示。\n\nAPI\n它不只供人下载模型，也提供在线托管和调用入口。\n\nPyTorch\n平台上大量模型代码与权重，天然连着 PyTorch 生态。", "relations": { "open-source-model": { "label": "分发…", "note": "很多开源模型先在这里发布。" }, "model-leaderboard": { "label": "承载…榜单", "note": "它常提供模型评测与排行入口。" }, "api": { "label": "提供…调用入口", "note": "除下载外，也能在线调用模型。" }, "pytorch": { "label": "连接…生态", "note": "大量模型与代码围绕该框架共享。" } } } } }, { "id": "human-in-the-loop", "name": "Human-in-the-loop", "layer": "L6", "era": "2020", "publishedAt": "2026-05-29T16:08:01.211Z", "relations": [ { "to": "agent" }, { "to": "alignment" }, { "to": "ai-regulation" }, { "to": "automation-job" } ], "track": "today", "i18n": { "en": { "fullName": "Human-in-the-loop", "factExplain": "A setup where people check or approve key AI decisions.", "humanExplain": "AI can be the self-checkout machine. Human-in-the-loop is the store worker who fixes “unexpected item” before everyone snaps.\n\nIt puts people into key AI decisions. You meet it in content review, health care, and risky approvals.", "humanExplainDisplay": "AI can be the ==self-checkout machine==.\nHuman-in-the-loop is the ==store worker==\nwho fixes “unexpected item”\nbefore everyone snaps.\n\nIt puts people into key AI decisions.\nYou meet it in content review,\nhealth care,\nand risky approvals.", "relationsNarrative": "Agent\nA more powerful Agent needs human checks at key steps.\n\nAlignment\nHuman-in-the-loop uses human judgment to keep AI on track.\n\nAI-regulation\nAI-regulation often requires human review in high-risk uses.\n\nAutomation-job\nHuman-in-the-loop makes automation more like teamwork than replacement.", "relations": { "agent": { "label": "adds human checks to …", "note": "Stronger Agents need people at key checkpoints." }, "alignment": { "label": "helps put … into practice", "note": "Alignment also needs real people to correct the AI." }, "ai-regulation": { "label": "is often required by …", "note": "High-risk AI often needs human review." }, "automation-job": { "label": "softens the shock of …", "note": "It makes automation feel more like teamwork." } } }, "zh": { "fullName": "人在回路中", "factExplain": "让人类参与 AI 决策或执行关键环节的方式。", "humanExplain": "AI 能一路往前冲，但方向盘别全撒手；关键路口还得人类坐副驾，免得它自信满满开进沟里。\n\n常用于审核、医疗和高风险审批，让流程更稳、更可控。", "humanExplainDisplay": "AI 能一路往前冲，\n但方向盘别全撒手；\n关键路口还得人类\n坐==副驾==，\n免得它自信满满\n开进==沟里==。\n\n常用于审核、医疗和高风险审批，\n让流程更稳、更可控。", "relationsNarrative": "Agent\nAgent 越能自主执行任务，越需要在关键节点加入人工确认和兜底。\n\nAlignment\nHuman-in-the-loop 是 Alignment 的现实做法之一，用人类判断持续纠偏。\n\nAi-regulation\n在高风险应用里，AI Regulation 常要求保留人工审核与最终决定权。\n\nAutomation-job\n它让自动化更像人机协作，能减缓岗位被一步到位替代的冲击。", "relations": { "agent": { "label": "给…加人工把关", "note": "Agent 越能干，越需要关键节点有人兜底。" }, "alignment": { "label": "作为…落地手段", "note": "对齐不只靠训练，也靠真人介入校正。" }, "ai-regulation": { "label": "常被…要求", "note": "高风险 AI 场景常要求保留人工复核。" }, "automation-job": { "label": "缓冲…冲击", "note": "它让自动化更像协作，不是立刻全替代。" } } } } }, { "id": "hyena-hierarchy", "name": "Hyena Hierarchy", "layer": "L6", "era": "2023", "publishedAt": "2026-06-21T14:27:16.865Z", "relations": [ { "to": "attention" }, { "to": "transformer" }, { "to": "sequence-modeling" }, { "to": "context-window" } ], "track": "ingest", "i18n": { "en": { "fullName": "Hyena Hierarchy", "factExplain": "A model design for reading long sequences without full attention.", "humanExplain": "Hyena is like a cafeteria line with a secret shortcut. Attention asks every kid about every kid. Hyena passes smart notes down the line instead.\n\nIt helps AI read long text faster. You meet it in research on long-context language models.", "humanExplainDisplay": "Hyena is like a cafeteria line with a ==secret shortcut==.\nAttention asks every kid about every kid.\nHyena passes smart notes down the line instead.\n\nIt helps AI read long text faster.\nYou meet it in research on long-context language models.", "relations": { "attention": { "label": "avoids full …", "note": "Hyena tries to skip full attention." }, "transformer": { "label": "competes with …", "note": "Hyena is an alternative model design." }, "sequence-modeling": { "label": "helps with …", "note": "Hyena is built for long sequences." }, "context-window": { "label": "aims at longer …", "note": "Hyena targets longer usable context." } } } } }, { "id": "hyperparameter-optimization", "name": "HPO", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "fine-tuning" }, { "to": "sgd" }, { "to": "optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Hyperparameter Optimization", "factExplain": "A method for searching training settings to make a model work better.", "humanExplain": "HPO is like making pancakes on a fussy stove. One heat notch makes fluffy pancakes, or a rubber frisbee.\n\nIt tries many training settings and keeps the best mix. You see it in training and fine-tuning. It costs time and compute.", "humanExplainDisplay": "HPO is like making pancakes\non a ==fussy stove==.\nOne heat notch makes ==fluffy pancakes==,\nor a rubber frisbee.\n\nIt tries many training settings\nand keeps the best mix.\nYou see it in training and fine-tuning.\nIt costs time and compute.", "relationsNarrative": "Parameter\nHPO tunes training settings, not weights the model learns.\n\nFine-tuning\nFine-tuning often uses HPO to find better training settings.\n\nSGD\nHPO often searches SGD settings like learning rate and momentum.\n\nOptimization\nHPO is an optimization problem about hyperparameter mixes.", "relations": { "parameter": { "label": "differs from …", "note": "It tunes training knobs, not learned weights." }, "fine-tuning": { "label": "is often used in …", "note": "Fine-tuning often uses it to find better training settings." }, "sgd": { "label": "tunes … settings", "note": "Learning rate and momentum are common search targets." }, "optimization": { "label": "is a kind of …", "note": "It searches for a better mix of training settings." } } }, "zh": { "fullName": "超参数优化（Hyperparameter Optimization）", "factExplain": "系统搜索超参数组合以提升模型效果的方法。", "humanExplain": "打工人做表格，字号差一号、行距多一点，观感就变了；超参也这样，得反复试到最顺手那版。\n\n常用于训练和微调，帮模型找到更稳更准的配置，但会更耗时耗算力。", "humanExplainDisplay": "打工人做表格，\n字号差一号、\n行距多一点，\n观感就变了；\n超参也这样，\n得反复试到\n==最顺手那版==。\n\n常用于训练和微调，\n帮模型找到更稳更准的配置，\n但会更耗时耗算力。", "relationsNarrative": "Parameter\n它调的是训练设置，不是模型自己学出来的权重。\n\nFine-tuning\n微调时常要靠它寻找更合适的训练配置。\n\nSGD\n学习率、动量这类 SGD 设置常由它来搜索。\n\nOptimization\n它属于优化问题，但优化对象是超参数组合。", "relations": { "parameter": { "label": "区别于…", "note": "它调训练旋钮，不直接学权重。" }, "fine-tuning": { "label": "常用于…", "note": "微调时常靠它找更合适配置。" }, "sgd": { "label": "会调…设置", "note": "学习率、动量等常是优化对象。" }, "optimization": { "label": "属于…范畴", "note": "本质是寻找更优训练配置。" } } } } }, { "id": "ibm-models", "name": "IBM Models", "layer": "L2", "era": "2023", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "foundation-model" }, { "to": "open-weights" }, { "to": "enterprise-ai-deployment" }, { "to": "on-premise-ai" } ], "track": "history", "i18n": { "en": { "fullName": "IBM Models (Granite, etc.)", "factExplain": "A family of business-focused foundation models released by IBM.", "humanExplain": "IBM Models are like the office grown-up with a clipboard. They check reports, code, and contracts before anyone starts a victory dance.\n\nYou meet them in company Q&A and code helpers. Teams can run them privately, with more control, trust, and clearer costs.", "humanExplainDisplay": "IBM Models are like the ==office grown-up==\nwith a ==clipboard==.\nThey check reports, code, and contracts\nbefore anyone starts a victory dance.\n\nYou meet them in company Q&A\nand code helpers.\nTeams can run them privately,\nwith more control, trust,\nand clearer costs.", "relationsNarrative": "Foundation-model\nIBM models are often delivered and used as foundation models.\n\nOpen weights\nSome Granite models offer open weights for easier self-deployment.\n\nEnterprise AI Deployment\nIBM models focus on controlled and trusted business rollout.\n\nOn-premise AI\nIBM models can support use inside a company’s private environment.", "relations": { "foundation-model": { "label": "belongs to … family", "note": "IBM models are often delivered as foundation models." }, "open-weights": { "label": "opens some …", "note": "Some Granite models provide open weights for self-deployment." }, "enterprise-ai-deployment": { "label": "serves …", "note": "IBM models focus on controlled, trusted business use." }, "on-premise-ai": { "label": "supports …", "note": "Companies can run them inside private environments." } } }, "zh": { "fullName": "IBM 模型（Granite 等）", "factExplain": "IBM 发布的企业向基础模型家族。", "humanExplain": "IBM 模型是公司里的合规管家：不抢镜，报表代码合同都先过一遍。\n\n用于企业问答、代码助手和私有部署，强调可控可信、成本可算。", "humanExplainDisplay": "IBM 模型是公司里的\n==合规管家==：\n不抢镜，\n报表代码合同都先过一遍。\n\n用于企业问答、代码助手，\n和私有部署，\n强调可控可信、成本可算。", "relationsNarrative": "Foundation-model\nIBM 模型多以基础模型形态交付和调用。\n\nOpen weights\n部分 Granite 模型开放权重，方便自部署。\n\nEnterprise AI Deployment\n它主打企业场景里的可控、可信落地。\n\nOn-premise AI\n它可支持企业在私有环境中部署使用。", "relations": { "foundation-model": { "label": "属于…家族", "note": "IBM 模型多以基础模型形态交付。" }, "open-weights": { "label": "开放部分…", "note": "部分 Granite 模型提供开放权重。" }, "enterprise-ai-deployment": { "label": "服务…", "note": "它主打企业场景的可控落地。" }, "on-premise-ai": { "label": "支持…", "note": "企业可在私有环境中部署使用。" } } } } }, { "id": "id3-algorithm", "name": "ID3", "layer": "L2", "era": "1986", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "decision-tree" }, { "to": "information-theory" }, { "to": "classification" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Iterative Dichotomiser 3", "factExplain": "A classification method for building decision trees with information gain.", "humanExplain": "ID3 is like a speed round of 20 Questions. It picks the question with the cleanest split first.\n\nIt turns a table into if-then sorting rules. You meet it in simple risk checks and surveys.", "humanExplainDisplay": "ID3 is like a ==speed round of 20 Questions==.\nIt picks the question with the ==cleanest split== first.\n\nIt turns a table into if-then sorting rules.\nYou meet it in simple risk checks and surveys.", "relationsNarrative": "Decision Tree\nID3 trains a decision tree by choosing each split question.\n\nInformation Theory\nID3 uses entropy and information gain to pick the best question.\n\nClassification\nID3 is used for classification. It puts examples into fixed classes.\n\nSupervised Learning\nID3 learns from labeled data. So it belongs to supervised learning.", "relations": { "decision-tree": { "label": "trains …", "note": "ID3 builds a readable decision tree." }, "information-theory": { "label": "uses … to split", "note": "Information gain comes from entropy math." }, "classification": { "label": "is used for …", "note": "It puts examples into fixed classes." }, "supervised-learning": { "label": "belongs to …", "note": "It needs training data with labels." } } }, "zh": { "fullName": "Iterative Dichotomiser 3，迭代二分器 3 算法", "factExplain": "一种用信息增益构建决策树的分类算法。", "humanExplain": "ID3像相亲角阿姨开问：先问房车户口，最快把人分到合适小板凳。\n\n把表格变成分类规则，常用于风控和问卷。", "humanExplainDisplay": "ID3像相亲角阿姨开问：\n先问==房车户口==，\n最快把人分到\n==合适小板凳==。\n\n把表格变成分类规则，\n常用于风控和问卷。", "relationsNarrative": "Decision Tree\nID3 是训练决策树的经典方法，负责选分裂问题。\n\nInformation Theory\n它用熵和信息增益，判断哪个问题最能分人。\n\nClassification\n它主要服务分类任务，把样本归到离散类别。\n\nSupervised Learning\n它从带标签数据学习规则，属于监督学习。", "relations": { "decision-tree": { "label": "训练…", "note": "ID3 生成一棵可解释的决策树。" }, "information-theory": { "label": "借用…衡量分裂", "note": "信息增益来自熵的计算。" }, "classification": { "label": "用于…", "note": "它把样本分到离散类别。" }, "supervised-learning": { "label": "属于…", "note": "训练数据需要已有标签。" } } } } }, { "id": "image-captioning", "name": "Image Captioning", "layer": "L4", "era": "2014", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "multimodal" }, { "to": "coco-dataset" }, { "to": "bleu" } ], "track": "history", "i18n": { "en": { "fullName": "Image Captioning", "factExplain": "A task where AI writes a plain sentence about an image.", "humanExplain": "Image captioning is your camera roll with a tiny sports announcer. It says, “Dog on sofa. Pizza in danger.”\n\nIt turns pictures into words. You meet it in photo search, screen readers, and online shops.", "humanExplainDisplay": "Image captioning is your camera roll\nwith a ==tiny sports announcer==.\nIt says,\n“==Dog on sofa. Pizza in danger.==”\n\nIt turns pictures into words.\nYou meet it in photo search,\nscreen readers,\nand online shops.", "relationsNarrative": "Computer Vision\nImage captioning says what the AI sees in plain words.\n\nMultimodal AI\nIt connects image understanding with text writing.\n\nCOCO Dataset\nCOCO provides many images with human-written captions.\n\nBLEU\nBLEU roughly checks how close a caption is to reference captions.", "relations": { "computer-vision": { "label": "turns … into words", "note": "It turns visual recognition into a text description." }, "multimodal": { "label": "links images and text", "note": "Image captioning is a classic image-and-text task." }, "coco-dataset": { "label": "trains and tests with …", "note": "COCO gives paired images and human captions." }, "bleu": { "label": "checks text with …", "note": "BLEU roughly compares a caption with reference captions." } } }, "zh": { "fullName": "图像描述", "factExplain": "让模型为图像生成自然语言描述的任务。", "humanExplain": "图像描述就是朋友圈配文嘴替：照片刚发，它立刻补上谁在哪儿干啥。\n\n让图片可搜索、可读屏，常用于相册、无障碍和电商。", "humanExplainDisplay": "图像描述就是\n==朋友圈配文嘴替==：\n照片刚发，\n它立刻补上==谁在哪儿干啥==。\n\n让图片可搜索、可读屏，\n常用于相册、无障碍\n和电商。", "relationsNarrative": "Computer Vision\n图像描述把“看见了什么”说成自然语言。\n\nMultimodal AI\n它是连接图像理解与语言生成的典型任务。\n\nCOCO Dataset\nCOCO 提供大量图片与人工描述配对。\n\nBLEU\nBLEU 常用于粗略评估描述与参考答案的相似度。", "relations": { "computer-vision": { "label": "扩展…输出", "note": "它把视觉识别变成文字说明。" }, "multimodal": { "label": "连接图像与文本", "note": "图像描述是典型图文任务。" }, "coco-dataset": { "label": "用…训练评测", "note": "COCO 提供图文配对数据。" }, "bleu": { "label": "用…评估文本", "note": "BLEU 可粗略衡量描述相似度。" } } } } }, { "id": "image-classification", "name": "Image Class.", "layer": "L4", "era": "2012", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "computer-vision" }, { "to": "cnn" }, { "to": "imagenet" } ], "track": "history", "i18n": { "en": { "fullName": "Image Classification", "factExplain": "A vision task that puts an image into a preset category.", "humanExplain": "Image classification is like your phone playing “What am I looking at?” It squints at a blurry photo and says cat, dog, or cupcake.\n\nYou meet it in photo search and factory checks. It sends each picture to the right bucket.", "humanExplainDisplay": "Image classification is like your phone playing\n==“What am I looking at?”==\nIt squints at a blurry photo\nand says ==cat, dog, or cupcake==.\n\nYou meet it in photo search and factory checks.\nIt sends each picture to the right bucket.", "relationsNarrative": "Classification\nImage classification is the classic classification task for pictures.\n\nComputer Vision\nImage classification is one of the basic tasks in computer vision.\n\nCNN\nCNNs were the main model for image classification for years.\n\nImageNet\nImageNet used many labeled images to push image classification forward.", "relations": { "classification": { "label": "is a kind of … task", "note": "Image classification is classification applied to pictures." }, "computer-vision": { "label": "is a basic task in …", "note": "It is one of the basic jobs in computer vision." }, "cnn": { "label": "was often built with …", "note": "CNNs were the main model for image classification for years." }, "imagenet": { "label": "was pushed forward by …", "note": "ImageNet helped image classification improve fast with many labeled images." } } }, "zh": { "fullName": "Image Classification（图像分类）", "factExplain": "把图像归入预设类别的视觉任务。", "humanExplain": "图像分类像家庭群认亲：照片糊成马赛克，也得喊出猫、狗还是二舅。\n\n它用于相册搜索、质检和影像分流，帮图片自动进对应筐。", "humanExplainDisplay": "图像分类像家庭群认亲：\n照片==糊成马赛克==，\n也得喊出\n==猫、狗还是二舅==。\n\n它用于相册搜索、质检\n和影像分流，\n帮图片自动进对应筐。", "relationsNarrative": "Classification\n图像分类是分类问题在视觉领域的典型形式。\n\nComputer Vision\n图像分类是计算机视觉最基础的任务之一。\n\nCNN\nCNN 曾长期是图像分类的主力模型。\n\nImageNet\nImageNet 用大规模标注推动图像分类突破。", "relations": { "classification": { "label": "属于…任务", "note": "图像分类是分类问题在视觉里的代表。" }, "computer-vision": { "label": "服务于…", "note": "它是计算机视觉最基础的任务之一。" }, "cnn": { "label": "常用…实现", "note": "CNN 曾长期是图像分类的主力模型。" }, "imagenet": { "label": "用…衡量进展", "note": "ImageNet 推动了图像分类的大爆发。" } } } } }, { "id": "image-inpainting", "name": "Image Inpainting", "layer": "L4", "era": "2000", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "ai-photo-editor" }, { "to": "computer-vision" } ], "track": "today", "i18n": { "en": { "fullName": "Image Inpainting", "factExplain": "A technique that fills missing parts of an image with believable new content.", "humanExplain": "Image inpainting is like fixing a scratched family photo. The new pixels must blend in, or Grandma gets two eyebrows.\n\nIt fixes old photos and removes unwanted stuff. It also helps change backgrounds and make AI edits.", "humanExplainDisplay": "Image inpainting is like fixing a ==scratched family photo==.\nThe new pixels must ==blend in==,\nor Grandma gets two eyebrows.\n\nIt fixes old photos\nand removes unwanted stuff.\nIt also helps change backgrounds\nand make AI edits.", "relationsNarrative": "Diffusion\nDiffusion is often used to create more natural filled areas.\n\nAI Photo Editor\nImage inpainting powers many one-tap erase and background change tools.\n\nComputer Vision\nImage inpainting is a Computer Vision task between understanding and generation.", "relations": { "diffusion": { "label": "often uses …", "note": "Diffusion helps the filled area look more natural." }, "ai-photo-editor": { "label": "powers …", "note": "Many one-tap erase tools use image inpainting." }, "computer-vision": { "label": "is part of …", "note": "It sits between seeing an image and making new image parts." } } }, "zh": { "fullName": "图像修补", "factExplain": "在图像缺失区域生成合理内容的技术。", "humanExplain": "图像修补像给旧墙补漆：坑要填平，花纹颜色得接到房东来了都看不出。\n\n用于修老照片、去杂物、换背景和生成式编辑。", "humanExplainDisplay": "图像修补像给==旧墙补漆==：\n坑要填平，\n花纹颜色得接到\n==房东来了都看不出==。\n\n用于修老照片、去杂物，\n换背景和生成式编辑。", "relationsNarrative": "Diffusion\n扩散模型常用于生成更自然的修补区域。\n\nAI Photo Editor\n图像修补支撑很多一键擦除和改背景功能。\n\nComputer Vision\n它是计算机视觉里连接理解与生成的任务。", "relations": { "diffusion": { "label": "常用…补图", "note": "扩散模型让补图区域更自然。" }, "ai-photo-editor": { "label": "支撑…", "note": "很多一键擦除功能靠它完成。" }, "computer-vision": { "label": "属于…任务", "note": "它是视觉理解与生成的交界任务。" } } } } }, { "id": "image-segmentation", "name": "Image Segmentation", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "object-detection" }, { "to": "image-classification" }, { "to": "u-net" } ], "track": "history", "i18n": { "en": { "fullName": "Image Segmentation", "factExplain": "A vision task that splits an image into regions, pixel by pixel.", "humanExplain": "Image Seg is like a fussy kid with crayons. It colors the dog, not the rug.\n\nYou meet it in self-driving cars and medical scans. Photo apps use it to cut clean edges.", "humanExplainDisplay": "Image Seg is like a ==fussy kid with crayons==.\nIt colors the ==dog, not the rug==.\n\nYou meet it in self-driving cars and medical scans.\nPhoto apps use it to cut clean edges.", "relationsNarrative": "Computer Vision\nImage segmentation is a basic task in Computer Vision.\n\nObject Detection\nObject Detection draws boxes. Image segmentation gives cleaner edges.\n\nImage Classification\nClassification labels the whole picture, while segmentation labels each pixel.\n\nU-Net\nU-Net is a classic network often used for image segmentation.", "relations": { "computer-vision": { "label": "is part of …", "note": "Image segmentation is a basic Computer Vision task." }, "object-detection": { "label": "is finer than …", "note": "Object Detection draws boxes. Segmentation traces the exact edges." }, "image-classification": { "label": "goes beyond …", "note": "Image Class. says what is there. Segmentation says where each pixel belongs." }, "u-net": { "label": "often uses …", "note": "U-Net is a classic network for medical image segmentation." } } }, "zh": { "fullName": "Image Segmentation（图像分割）", "factExplain": "将图像按像素划分为不同区域的视觉任务。", "humanExplain": "图像分割像修图师抠头发丝：猫归猫、人归人，边边角角都不糊弄。\n\n用于自动驾驶、医学影像和修图，帮机器看清边界。", "humanExplainDisplay": "图像分割像修图师\n==抠头发丝==：\n猫归猫、人归人，\n边边角角都不糊弄。\n\n用于自动驾驶、医学影像，\n和修图，\n帮机器看清边界。", "relationsNarrative": "Computer Vision\n图像分割是计算机视觉里的基础任务。\n\nObject Detection\n检测画框，分割给出更精确的边界。\n\nImage Classification\n分类判断整张图，分割判断每个像素。\n\nU-Net\nU-Net 是图像分割常用的经典网络。", "relations": { "computer-vision": { "label": "属于…", "note": "图像分割是计算机视觉的基础任务。" }, "object-detection": { "label": "比…更精细", "note": "检测找框，分割描出精确边界。" }, "image-classification": { "label": "细化…", "note": "分类说有什么，分割说在哪儿。" }, "u-net": { "label": "常用…实现", "note": "U-Net 是医学分割的经典架构。" } } } } }, { "id": "imagenet", "name": "ImageNet", "layer": "L4", "era": "2009", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "alexnet" }, { "to": "cnn" }, { "to": "computer-vision" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "ImageNet", "factExplain": "A large image dataset and contest benchmark for visual recognition.", "humanExplain": "ImageNet is a giant spelling bee for robot eyes. The host flashes photos, and each AI shouts, “Dog!” or “Toaster!”\n\nPeople use it to train image models and compare them. For years, it was the classic scoreboard for Computer Vision.", "humanExplainDisplay": "ImageNet is a ==giant spelling bee==\nfor robot eyes.\nThe host flashes photos,\nand each AI shouts,\n==“Dog!” or “Toaster!”==\n\nPeople use it to train image models\nand compare them.\nFor years,\nit was the classic scoreboard\nfor Computer Vision.", "relationsNarrative": "AlexNet\nAlexNet won big on ImageNet and made deep learning famous.\n\nCNN\nImageNet was a classic training and testing ground for CNNs.\n\nComputer Vision\nImageNet was one of the best-known public tests in early Computer Vision.\n\nClassification\nImageNet is best known as a picture Classification test.", "relations": { "alexnet": { "label": "made … famous", "note": "AlexNet won big on ImageNet and made deep learning famous." }, "cnn": { "label": "trained many …", "note": "ImageNet was a key practice field for CNNs." }, "computer-vision": { "label": "benchmarks …", "note": "For years, ImageNet set a main test for Computer Vision." }, "classification": { "label": "is used for …", "note": "Its classic task is sorting pictures into labels." } } }, "zh": { "fullName": "ImageNet 图像数据集", "factExplain": "一个大规模视觉识别数据集与竞赛基准。", "humanExplain": "ImageNet 像武林大会擂台，谁家识图功夫真够硬，上去打几轮，名号立马传遍江湖。\n\n它用来训练和比较识图模型，长期是视觉领域检验实力的经典基准。", "humanExplainDisplay": "ImageNet 像武林大会==擂台==，\n谁家识图功夫真够硬，\n上去打几轮，\n==名号立马传遍江湖==。\n\n它用来训练和比较识图模型，\n长期是视觉领域检验实力的经典基准。", "relationsNarrative": "AlexNet\nAlexNet 在 ImageNet 竞赛上大幅领先，带火深度学习。\n\nCNN\nImageNet 长期是 CNN 训练和比拼效果的经典场地。\n\nComputer-vision\n它是计算机视觉早期最有代表性的公共基准之一。\n\nClassification\nImageNet 最典型的任务，就是图片分类。", "relations": { "alexnet": { "label": "让…一战成名", "note": "AlexNet 在它的竞赛上掀起深度学习浪潮。" }, "cnn": { "label": "常用…刷榜", "note": "它长期是卷积网络的重要练兵场。" }, "computer-vision": { "label": "是…经典基准", "note": "它长期定义了视觉识别的主流评测方式。" }, "classification": { "label": "常被用于…任务", "note": "它最经典的任务就是给图片分门别类。" } } } } }, { "id": "imitation-learning", "name": "Imitation Learning", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "supervised-learning" }, { "to": "robotics" } ], "track": "today", "i18n": { "en": { "fullName": "Imitation Learning", "factExplain": "A way for AI to learn by copying expert examples.", "humanExplain": "Imitation Learning is like learning a TikTok dance from your older cousin. You watch, copy, and try not to copy the weird face too.\n\nRobots and self-driving cars use it to learn faster. It can also copy bad habits from the teacher.", "humanExplainDisplay": "Imitation Learning is like learning a ==TikTok dance==\nfrom your older cousin.\nYou watch, copy,\nand try not to copy the ==weird face== too.\n\nRobots and self-driving cars use it\nto learn faster.\nIt can also copy bad habits\nfrom the teacher.", "relationsNarrative": "RL\nImitation Learning uses expert examples to cut RL trial and error.\n\nSupervised Learning\nBehavior cloning learns by treating expert actions as labels.\n\nRobotics\nRobots often learn actions from human demos.", "relations": { "reinforcement-learning": { "label": "cuts down … trial and error", "note": "Expert examples help RL avoid many wrong tries." }, "supervised-learning": { "label": "trains in a … style", "note": "It treats the expert’s actions like labels to learn from." }, "robotics": { "label": "teaches … actions", "note": "Robots often learn skills by watching human demos." } } }, "zh": { "fullName": "模仿学习", "factExplain": "通过专家示范数据学习行为策略的方法。", "humanExplain": "模仿学习像学包饺子：别先背面皮力学，看妈妈捏，跟着捏。\n\n用于机器人和自动驾驶，学得快，也会抄坏习惯。", "humanExplainDisplay": "模仿学习像==学包饺子==：\n别先背面皮力学，\n看妈妈捏，\n==跟着捏==。\n\n用于机器人和自动驾驶，\n学得快，\n也会抄坏习惯。", "relationsNarrative": "Reinforcement Learning\n模仿学习常用示范减少 RL 的试错成本。\n\nSupervised Learning\n行为克隆把示范动作当标签来学。\n\nRobotics\n机器人常靠人类演示学会动作。", "relations": { "reinforcement-learning": { "label": "减少…试错", "note": "专家示范能让 RL 少走弯路。" }, "supervised-learning": { "label": "借…形式训练", "note": "把示范动作当成监督标签。" }, "robotics": { "label": "教…学动作", "note": "机器人常从人类演示学技能。" } } } } }, { "id": "in-browser-ai-ai", "name": "In-browser AI", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "api" }, { "to": "local-llm" }, { "to": "ai-device-ai" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "In-browser AI", "factExplain": "A way to run or use AI right inside a web browser.", "humanExplain": "In-browser AI is like a mini fridge in your browser. Need a quick AI snack? Do not install the whole kitchen.\n\nYou meet it in web helpers and online docs. It can draft small text or handle a file. It opens fast, but your device sets the speed.", "humanExplainDisplay": "In-browser AI is like a ==mini fridge==\nin your browser.\nNeed a quick AI snack?\nDo not install the ==whole kitchen==.\n\nYou meet it in web helpers and online docs.\nIt can draft small text or handle a file.\nIt opens fast,\nbut your device sets the speed.", "relationsNarrative": "API\nIn-browser AI often uses an API to call a model on a server.\n\nLocal-LLM\nIt can also run a small Local-LLM inside the browser.\n\nAI device\nIt is one way AI device power shows up for users.\n\nData-privacy\nIf data stays on your device, privacy risk is usually lower.", "relations": { "api": { "label": "often connects through …", "note": "Many web AI tools use an API to reach a model on a server." }, "local-llm": { "label": "can run …", "note": "A browser can run a small Local-LLM on your own device." }, "ai-device-ai": { "label": "is a form of …", "note": "It puts AI power on the device you already use." }, "data-privacy": { "label": "changes … limits", "note": "If data stays local, fewer private details leave your device." } } }, "zh": { "fullName": "浏览器内 AI", "factExplain": "直接在浏览器中运行或调用 AI 能力的方式。", "humanExplain": "它更像把 AI 搬进便利店窗口，路过网页就能顺手用，不用先扛回一整套家电再开张。\n\n常见于网页助手、文档处理、轻生成；优点是即开即用，性能受设备影响。", "humanExplainDisplay": "它更像把 AI 搬进\n==便利店窗口==，\n路过网页就能顺手用，\n不用先扛回==一整套家电==再开张。\n\n常见于网页助手、\n文档处理、轻生成；\n优点是即开即用，\n性能受设备影响。", "relationsNarrative": "API\n很多浏览器内 AI 实际通过 API 调用后端模型。\n\nLocal-LLM\n它也能在本地浏览器中运行轻量模型。\n\nAI Device\n它是端侧 AI 的常见落地入口之一。\n\nData Privacy\n数据若留在本地处理，隐私风险通常更低。", "relations": { "api": { "label": "常通过…接模型", "note": "很多网页 AI 靠 API 连后端模型。" }, "local-llm": { "label": "可作为…入口", "note": "也能在本地浏览器里跑轻量模型。" }, "ai-device-ai": { "label": "属于…落地形态", "note": "把 AI 能力直接塞进用户终端。" }, "data-privacy": { "label": "影响…边界", "note": "本地处理可减少敏感数据外传。" } } } } }, { "id": "in-context-learning", "name": "In-Context Learning", "layer": "L2", "era": "2020", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "prompt" }, { "to": "context-window" }, { "to": "instruction-tuning" } ], "track": "history", "i18n": { "en": { "fullName": "In-context Learning", "factExplain": "A way for a model to learn a task from examples in the prompt.", "humanExplain": "In-context learning is like seeing two perfect group-chat replies first. You copy the pattern and do not panic-type “lol help.”\n\nYou meet it in few-shot prompts and format copying. The model learns on the spot, so you skip retraining.", "humanExplainDisplay": "In-context learning is like seeing\n==two perfect group-chat replies== first.\nYou ==copy the pattern==\nand do not panic-type “lol help.”\n\nYou meet it in few-shot prompts\nand format copying.\nThe model learns on the spot,\nso you skip retraining.", "relationsNarrative": "Prompt\nIn-context learning is usually triggered by instructions and examples in the prompt.\n\nContext-window\nIn-context learning can only use examples that fit in the context-window.\n\nInstruction Tuning\nIn-context learning picks up the pattern on the spot. Instruction Tuning practices it ahead of time.", "relations": { "prompt": { "label": "triggered by …", "note": "The examples are usually written inside the prompt." }, "context-window": { "label": "limited by …", "note": "The window size controls how many examples can fit." }, "instruction-tuning": { "label": "works with …", "note": "In-context learning learns on the spot. Instruction tuning practices ahead of time." } } }, "zh": { "fullName": "上下文学习", "factExplain": "模型根据上下文示例临时学会任务的方法。", "humanExplain": "上下文学习像相亲前先看两位成功案例：不用回炉进修，照着话术和节奏，当场也能接得上。\n\n常用于 few-shot 提示、格式模仿等，省得重新训练。", "humanExplainDisplay": "上下文学习像相亲前\n先看两位==成功案例==：\n不用回炉进修，\n照着话术和节奏，\n当场也能==接得上==。\n\n常用于 few-shot 提示、\n格式模仿等，\n省得重新训练。", "relationsNarrative": "Prompt\n它通常靠提示词里的说明和示例被触发。\n\nContext-window\n它能参考多少示例，受上下文窗口大小限制。\n\nInstruction-tuning\n它是临场学套路，指令微调则是提前练熟。", "relations": { "prompt": { "label": "靠…触发", "note": "示例通常就写在提示词里。" }, "context-window": { "label": "受…限制", "note": "能塞多少例子取决于窗口大小。" }, "instruction-tuning": { "label": "常与…互补", "note": "前者临时学，后者提前练。" } } } } }, { "id": "independent-component-analysis", "name": "ICA", "layer": "L2", "era": "1994", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "unsupervised-learning" }, { "to": "principal-component-analysis" }, { "to": "dimensionality-reduction" }, { "to": "latent-variable-model" } ], "track": "history", "seo": { "zh": { "title": "独立成分分析 是什么?KTV合唱拆轨,一文看懂 — AI Rookies", "description": "把混合信号分解为相互独立成分的方法。人话版一分钟讲明白。" }, "en": { "title": "What Is ICA? Each Sound Its Own Lane", "description": "A method that splits mixed signals into independent hidden parts. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Independent Component Analysis", "factExplain": "A method that splits mixed signals into independent hidden parts.", "humanExplain": "ICA is like hearing a call in a wild kitchen. The blender screams, the dog barks, and Dad sings off-key. ICA tries to give each sound its own lane.\n\nIt finds hidden sources in mixed data. You meet it in voice cleanup. You also meet it with brain waves and images.", "humanExplainDisplay": "ICA is like hearing a call\nin a ==wild kitchen==.\nThe blender screams,\nthe dog barks,\nand Dad sings off-key.\nICA tries to give each sound\n==its own lane==.\n\nIt finds hidden sources\nin mixed data.\nYou meet it in voice cleanup.\nYou also meet it with brain waves\nand images.", "relationsNarrative": "Unsupervised Learning\nICA uses no labels and finds hidden structure in the data.\n\nPCA\nPCA looks for uncorrelated parts. ICA looks for independent parts.\n\nDim. Reduction\nICA can turn mixed signals into clearer feature form.\n\nLatent Model\nICA treats independent parts as hidden variables behind the data.", "relations": { "unsupervised-learning": { "label": "belongs to …", "note": "ICA uses no labels. It finds structure in mixed data." }, "principal-component-analysis": { "label": "separates sources better than …", "note": "PCA looks for uncorrelated parts. ICA looks for independent parts." }, "dimensionality-reduction": { "label": "can be used for …", "note": "ICA can turn mixed signals into clearer features." }, "latent-variable-model": { "label": "finds hidden factors in …", "note": "Independent parts can act like hidden variables behind what we observe." } } }, "zh": { "fullName": "独立成分分析", "factExplain": "把混合信号分解为相互独立成分的方法。", "humanExplain": "ICA像在KTV包厢听合唱：麦霸、跑调和伴奏再吵，也能一人一轨拆开。\n\n用于语音分离、脑电和图像，帮人从混合数据找源头。", "humanExplainDisplay": "ICA像在KTV包厢\n听合唱：\n麦霸跑调伴奏==再吵==，\n也能==一人一轨拆开==。\n\n用于语音分离、脑电和图像，\n帮人从混合数据找源头。", "relationsNarrative": "Unsupervised Learning\nICA 不用标签，而是从数据本身拆出隐藏结构。\n\nPCA\nPCA 追求不相关，ICA 更进一步追求统计独立。\n\nDimensionality Reduction\nICA 可把混合信号压成更清晰的特征表示。\n\nLatent Model\nICA 找到的独立成分可看作背后的隐变量。", "relations": { "unsupervised-learning": { "label": "属于…", "note": "它不用标签，从混合数据里找结构。" }, "principal-component-analysis": { "label": "比…更会拆源", "note": "PCA看不相关，ICA追求统计独立。" }, "dimensionality-reduction": { "label": "可用于…", "note": "它常把混合信号变成清晰特征。" }, "latent-variable-model": { "label": "寻找…中的隐因子", "note": "独立成分像藏在观测背后的变量。" } } } } }, { "id": "inductive-bias", "name": "Inductive Bias", "layer": "L2", "era": "1980", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "transformer" }, { "to": "bias-variance-tradeoff" }, { "to": "statistical-learning-theory" } ], "track": "history", "i18n": { "en": { "fullName": "Inductive Bias", "factExplain": "The built-in guesses a model uses before it learns from data.", "humanExplain": "It is like a detective at a missing-cookie scene. He checks the dog first, not Grandma.\n\nIt shapes what a model learns fastest. It matters most with little data and in model design.", "humanExplainDisplay": "It is like a ==detective==\nat a missing-cookie scene.\nHe checks ==the dog first==,\nnot Grandma.\n\nIt shapes what a model learns fastest.\nIt matters most with little data\nand in model design.", "relationsNarrative": "CNN\nInductive bias in CNNs favors nearby patterns and the same pattern in new spots.\n\nTransformer\nA Transformer has its own design guesses built into it.\n\nBias-Variance Tradeoff\nStronger inductive bias can lower variance but raise bias.\n\nSLT\nSLT studies how inductive bias helps models work on new data.", "relations": { "cnn": { "label": "shows up in …", "note": "CNNs prefer nearby patterns and the same pattern in new spots." }, "transformer": { "label": "sets …'s preferences", "note": "Each design builds different starting guesses into the model." }, "bias-variance-tradeoff": { "label": "shifts …", "note": "Stronger bias can lower variance but raise bias." }, "statistical-learning-theory": { "label": "is studied by …", "note": "SLT asks how models learn from samples and still work on new data." } } }, "zh": { "fullName": "归纳偏置", "factExplain": "模型先天带着的假设和偏好。", "humanExplain": "像老中医一搭脉，心里先有路数：还没开方，就默认有些症状更值得盯紧。\n\n它会影响模型更容易学到什么，在小样本和架构设计里尤其关键。", "humanExplainDisplay": "像老中医一搭脉，\n心里先有==路数==：\n还没开方，\n就默认有些症状\n更值得==盯紧==。\n\n它会影响模型\n更容易学到什么，\n在小样本和架构设计里\n尤其关键。", "relationsNarrative": "CNN\n卷积网络把局部模式和平移不变性写进了结构里。\n\nTransformer\n不同架构的设计，本质上就是不同归纳偏置。\n\nBias-variance-tradeoff\n归纳偏置越强，通常越容易降低方差但带来偏差。\n\nStatistical-learning-theory\n它是统计学习理论讨论泛化能力时的核心问题。", "relations": { "cnn": { "label": "体现在…里", "note": "卷积结构自带对局部与平移的偏好。" }, "transformer": { "label": "决定…偏好", "note": "不同架构把不同先验写进模型里。" }, "bias-variance-tradeoff": { "label": "影响…平衡", "note": "先验越强，越可能降方差增偏差。" }, "statistical-learning-theory": { "label": "属于…关心的问题", "note": "它解释模型为何能从样本推广到新数据。" } } } } }, { "id": "inference", "name": "Inference", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-05-23T08:40:00Z", "relations": [ { "to": "llm" }, { "to": "token" }, { "to": "gpu" }, { "to": "api" } ], "i18n": { "en": { "fullName": "Inference", "factExplain": "The run step where a trained model takes input and makes output.", "humanExplain": "Training is teaching a barista the menu. Inference is your latte getting made after you order.\n\nEvery chat reply, image, or model call runs inference. It spends time and computer power.", "humanExplainDisplay": "Training is teaching a ==barista== the menu.\nInference is your ==latte getting made== after you order.\n\nEvery chat reply, image, or model call runs inference.\nIt spends time and computer power.", "relationsNarrative": "LLM\nInference is when an LLM uses its learned skills to answer a request.\n\nToken\nInference reads input as Tokens and writes output step by step.\n\nGPU\nGPU power affects the speed and cost of inference.\n\nAPI\nAn API wraps inference so apps can call it.", "relations": { "llm": { "label": "runs the …", "note": "Inference is when an LLM uses its learned skills to answer." }, "token": { "label": "processes …", "note": "Inference reads and writes text as Tokens." }, "gpu": { "label": "depends on … for speed", "note": "A faster GPU can make inference quicker and cheaper." }, "api": { "label": "is often offered through an …", "note": "An API lets apps call inference like a service." } } }, "zh": { "fullName": "推理（运行）", "factExplain": "模型接收输入后生成输出的运行过程。", "humanExplain": "推理像餐厅真正开始出菜，训练是备料，你点单后厨房才开火。\n\n每次聊天、生成图片、调用模型，背后都在消耗算力和时间。", "humanExplainDisplay": "推理像餐厅真正\n==开始出菜==。\n训练是备料，你点单后才开火。\n\n每次聊天、生成图片、调用模型，\n都在消耗算力和时间。\n慢不是装深沉，是真在算。", "relationsNarrative": "LLM\nInference 是 LLM 利用已学能力响应请求的阶段。\n\nToken\nInference 会按 Token 读取输入，并逐步生成输出。\n\nGPU\nGPU 性能直接影响 Inference 的速度和成本。\n\nAPI\nAPI 将 Inference 封装成应用可调用的能力。", "relations": { "llm": { "label": "运行…" }, "token": { "label": "处理…" }, "gpu": { "label": "依赖…加速" }, "api": { "label": "常以…对外提供" } } } } }, { "id": "infini-attention-transformer", "name": "Infini-attention Transformer", "layer": "L3", "era": "2024", "publishedAt": "2026-06-21T14:27:25.107Z", "relations": [ { "to": "attention" }, { "to": "transformer" }, { "to": "context-window" }, { "to": "context-compression" } ], "track": "ingest", "i18n": { "en": { "fullName": "Infini-Transformer", "factExplain": "A Transformer design with compressed memory for very long context.", "humanExplain": "Infini-attention is a Transformer with a tiny desk and a magic junk drawer. The desk holds today’s page. The drawer keeps squished notes from the whole book.\n\nIt helps AI read huge text without keeping every bit in full. You will mostly see it in long-context AI research.", "humanExplainDisplay": "Infini-attention is a Transformer with a tiny desk and a ==magic junk drawer==.\nThe desk holds today’s page.\nThe drawer keeps squished notes from the whole book.\n\nIt helps AI read huge text without keeping every bit in full.\nYou will mostly see it in long-context AI research.", "relations": { "attention": { "label": "adds memory to …", "note": "It adds a long memory path to attention." }, "transformer": { "label": "upgrades …", "note": "A Transformer built for huge context." }, "context-window": { "label": "pushes past …", "note": "It aims to stretch how much AI can read." }, "context-compression": { "label": "uses …", "note": "Old text becomes small memory notes." } } } } }, { "id": "information-bottleneck", "name": "Information Bottleneck", "layer": "L2", "era": "1999", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "information-theory" }, { "to": "representation-learning" }, { "to": "dimensionality-reduction" }, { "to": "kullback-leibler-divergence" } ], "track": "history", "i18n": { "en": { "fullName": "Information Bottleneck", "factExplain": "A rule for keeping task clues while squeezing data into a smaller form.", "humanExplain": "Information Bottleneck is packing for a weekend with one tiny backpack. Socks stay. Your rock collection stays home.\n\nYou meet it in representation learning and compression. It keeps task clues and drops noise.", "humanExplainDisplay": "Information Bottleneck is ==packing for a weekend==\nwith one tiny backpack.\n==Socks stay==.\nYour rock collection stays home.\n\nYou meet it in representation learning\nand compression.\nIt keeps task clues\nand drops noise.", "relationsNarrative": "Information Theory\nInformation Bottleneck uses information amount to choose what to keep and what to drop.\n\nRepresentation Learning\nIt guides models to learn smaller representations with more task clues.\n\nDim. Reduction\nBoth compress data, but Information Bottleneck cares more about task relevance.\n\nKL Divergence\nKL often measures the information gap before and after compression.", "relations": { "information-theory": { "label": "comes from …", "note": "It uses information amount to decide what stays and goes." }, "representation-learning": { "label": "guides …", "note": "A good representation keeps task clues and carries less noise." }, "dimensionality-reduction": { "label": "compresses like …", "note": "Both make data smaller, but this cares more about useful clues." }, "kullback-leibler-divergence": { "label": "often uses … as a constraint", "note": "KL can measure the information gap before and after compression." } } }, "zh": { "fullName": "信息瓶颈", "factExplain": "用压缩表示保留与目标相关信息的原则。", "humanExplain": "信息瓶颈像高考小抄只够一掌大：考点留下，闲话统统删掉。\n\n用于表示学习和压缩，帮助模型只保留任务线索。", "humanExplainDisplay": "信息瓶颈像高考小抄\n只够==一掌大==：\n==考点留下==，\n闲话统统删掉。\n\n用于表示学习和压缩，\n帮助模型只保留\n任务线索。", "relationsNarrative": "Information Theory\n信息瓶颈用信息量刻画哪些该保留、哪些该丢弃。\n\nRepresentation Learning\n它指导模型学到更紧凑、与任务更相关的表示。\n\nDimensionality Reduction\n二者都压缩数据，但信息瓶颈更关心任务相关性。\n\nKL Divergence\nKL 常用于衡量压缩前后分布的信息差。", "relations": { "information-theory": { "label": "源自…", "note": "它用信息量衡量保留与丢弃。" }, "representation-learning": { "label": "指导…", "note": "好表示要少带噪声，多留任务线索。" }, "dimensionality-reduction": { "label": "类似…压缩", "note": "都在变小，但它更强调相关信息。" }, "kullback-leibler-divergence": { "label": "常用…约束", "note": "KL 可度量压缩前后的信息差。" } } } } }, { "id": "information-extraction", "name": "IE", "layer": "L4", "era": "1980s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "document-parsing" }, { "to": "sequence-labeling" }, { "to": "knowledge-graph" } ], "track": "history", "i18n": { "en": { "fullName": "Information Extraction", "factExplain": "A technique that pulls clear fields from messy text or documents.", "humanExplain": "Information Extraction is like cleaning a messy backpack. It grabs the permission slip and lunch money, then leaves the mystery crumbs alone.\n\nAt work, it reads messy documents and online chatter. It pulls key fields into tables for search and analysis.", "humanExplainDisplay": "Information Extraction is like\n==cleaning a messy backpack==.\nIt grabs the ==permission slip and lunch money==,\nthen leaves the mystery crumbs alone.\n\nAt work, it reads messy documents and online chatter.\nIt pulls key fields into tables\nfor search and analysis.", "relationsNarrative": "NLP\nInformation Extraction is a classic NLP task for turning text into structure.\n\nDocument parsing\nDocument parsing breaks the page first. Information Extraction then grabs the fields.\n\nSeq Labeling\nSeq Labeling helps mark useful pieces in the text.\n\nKnowledge Graph\nExtracted entities and links can become material for a Knowledge Graph.", "relations": { "natural-language-processing": { "label": "is a … task", "note": "It is a classic NLP task for organizing text." }, "document-parsing": { "label": "comes after …", "note": "Document parsing breaks the page before IE pulls fields." }, "sequence-labeling": { "label": "often uses …", "note": "Seq Labeling helps find entities and fields in text." }, "knowledge-graph": { "label": "feeds …", "note": "Extracted entities and links can go into a Knowledge Graph." } } }, "zh": { "fullName": "信息抽取", "factExplain": "从非结构化内容中抽取结构化信息的技术。", "humanExplain": "信息抽取像麻辣烫夹菜：一锅乱炖里，只挑丸子青菜，按格放盘不端汤。\n\n它把合同、简历、舆情中的字段抽成表，便于检索分析。", "humanExplainDisplay": "信息抽取像麻辣烫夹菜：\n一锅乱炖里，\n==只挑丸子青菜==，\n按格放盘==不端汤==。\n\n它把合同、简历、舆情\n中的字段抽成表，\n便于检索分析。", "relationsNarrative": "NLP\n信息抽取是 NLP 中把文本变结构的经典任务。\n\nDocument parsing\n文档解析先拆开版面，信息抽取再抓字段。\n\nSequence Labeling\n序列标注常用于识别实体、时间和字段。\n\nKnowledge Graph\n抽出的实体和关系可成为知识图谱的材料。", "relations": { "natural-language-processing": { "label": "属于…任务", "note": "它是 NLP 的经典信息整理任务。" }, "document-parsing": { "label": "接在…之后", "note": "先拆开文档，才能稳定抽字段。" }, "sequence-labeling": { "label": "常用…识别片段", "note": "序列标注常用来找实体和字段。" }, "knowledge-graph": { "label": "为…供料", "note": "抽出的实体关系可写入知识图谱。" } } } } }, { "id": "information-retrieval", "name": "IR", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "vector-search" }, { "to": "rag" }, { "to": "embedding" }, { "to": "vector-db" } ], "track": "history", "i18n": { "en": { "fullName": "Information Retrieval", "factExplain": "Technology that finds related content inside a huge pile of information.", "humanExplain": "Information retrieval is like searching a messy group chat for the pizza plan. It skips the memes and finds the actual address.\n\nYou meet it in search bars, knowledge bases, and Q&A tools. It finds related material first.", "humanExplainDisplay": "Information retrieval is like searching a ==messy group chat==\nfor the pizza plan.\nIt skips the memes\nand finds the ==actual address==.\n\nYou meet it in search bars,\nknowledge bases,\nand Q&A tools.\nIt finds related material first.", "relationsNarrative": "Vector search\nVector search is a key modern way to do information retrieval.\n\nRAG\nRAG uses information retrieval to find sources before the model writes.\n\nEmbedding\nEmbedding turns content into vectors, so search can match meaning.\n\nVector-db\nA vector database often stores and searches the retrieval index.", "relations": { "vector-search": { "label": "includes … methods", "note": "Vector search is one modern way to do information retrieval." }, "rag": { "label": "finds sources for …", "note": "RAG uses it to fetch related content before answering." }, "embedding": { "label": "uses … for meaning", "note": "Embedding makes similar content easier to find." }, "vector-db": { "label": "often uses …", "note": "A vector database often stores the search index." } } }, "zh": { "fullName": "信息检索（Information Retrieval）", "factExplain": "从大量信息中找到相关内容的技术。", "humanExplain": "信息检索像图书馆找资料：你报个题目，老师傅不陪你翻全馆，先把最对路的那几本拎出来。\n\n它常见于搜索、知识库和问答，先从海量内容里找相关材料。", "humanExplainDisplay": "信息检索像图书馆找资料：\n你报个题目，\n老师傅不陪你翻全馆，\n先把==最对路==的那几本==拎出来==。\n\n它常见于搜索、\n知识库和问答，\n先从海量内容里找相关材料。", "relationsNarrative": "Vector Search\n向量检索是信息检索的一种重要现代做法。\n\nRAG\nRAG 先用信息检索找资料，再交给模型生成。\n\nEmbedding\nEmbedding 把内容变成向量，便于按语义检索。\n\nVector DB\n向量数据库常用来存放和查询检索索引。", "relations": { "vector-search": { "label": "包含…方法", "note": "向量检索是它的现代实现之一。" }, "rag": { "label": "为…先找资料", "note": "RAG 先靠它找相关内容再作答。" }, "embedding": { "label": "常借助…表示", "note": "Embedding 让相似内容更易被找到。" }, "vector-db": { "label": "常配合…落地", "note": "向量数据库常承载检索索引。" } } } } }, { "id": "information-theory", "name": "Information Theory", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "cross-entropy-loss" }, { "to": "kullback-leibler-divergence" }, { "to": "token" }, { "to": "statistical-learning-theory" } ], "track": "history", "i18n": { "en": { "fullName": "Information Theory", "factExplain": "The theory of measuring, coding, and sending information.", "humanExplain": "Info Theory is like packing a lunchbox for a bumpy bus ride. You want it snug, safe, and still edible at noon.\n\nIt tells us how to shrink messages and send them without mangling them. In AI, it shapes entropy and training loss.", "humanExplainDisplay": "Info Theory is like packing a ==lunchbox==\nfor a ==bumpy bus ride==.\nYou want it snug,\nsafe,\nand still edible at noon.\n\nIt tells us how to shrink messages\nand send them without mangling them.\nIn AI,\nit shapes entropy and training loss.", "relationsNarrative": "Cross-Entropy Loss\nCross-Entropy Loss brings entropy from Information Theory into model training.\n\nKL Divergence\nKL Divergence is a classic Information Theory tool for comparing distributions.\n\nToken\nInformation Theory can measure how much information one Token carries.\n\nSLT\nInformation Theory covers information, and SLT covers learning and generalization.", "relations": { "cross-entropy-loss": { "label": "lays the base for …", "note": "Cross-entropy comes straight from information and entropy." }, "kullback-leibler-divergence": { "label": "defines … measure", "note": "KL Divergence measures how far two distributions differ." }, "token": { "label": "measures info in …", "note": "Information Theory can put a number on one token's information." }, "statistical-learning-theory": { "label": "pairs with …", "note": "Information Theory tracks information. SLT tracks learning and generalization." } } }, "zh": { "fullName": "信息论（Information Theory）", "factExplain": "研究信息度量、编码与传输的理论。", "humanExplain": "信息论像寄快递：既想箱子塞得紧省运费，又怕路上磕坏，还得收件人一拆就对货。\n\n它决定压缩和通信如何更省、更稳，也影响模型里的熵和损失设计。", "humanExplainDisplay": "信息论像寄快递：\n既想箱子==塞得紧==省运费，\n又怕路上磕坏，\n还得收件人一拆就==对货==。\n\n它决定压缩和通信如何更省、\n更稳，\n也影响模型里的熵和损失设计。", "relationsNarrative": "Cross-Entropy Loss\n交叉熵把信息论里的熵搬进了模型训练。\n\nKL Divergence\nKL 散度是信息论里比较分布差异的经典工具。\n\nToken\n信息论可描述一个符号到底有多“值钱”。\n\nStatistical Learning Theory\n它关注信息；SLT 更关注学习与泛化。", "relations": { "cross-entropy-loss": { "label": "奠定…基础", "note": "交叉熵直接来自信息量与熵概念。" }, "kullback-leibler-divergence": { "label": "定义…度量", "note": "KL 散度衡量两个分布差多远。" }, "token": { "label": "帮助度量…信息量", "note": "单个符号能携带多少信息可被量化。" }, "statistical-learning-theory": { "label": "补足…视角", "note": "一个管信息，一个管泛化与学习。" } } } } }, { "id": "instruction-tuning", "name": "Instruction Tuning", "layer": "L2", "era": "2021", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "fine-tuning" }, { "to": "rlhf" }, { "to": "base-model" } ], "track": "history", "i18n": { "en": { "fullName": "Instruction Tuning", "factExplain": "A training step for teaching a model to follow requests better.", "humanExplain": "Instruction tuning is like training a golden retriever with homework. You say “summarize this,” and it learns not to bring you a tennis ball.\n\nIt turns a base model into a better helper. You meet it inside chatbots and AI assistants.", "humanExplainDisplay": "Instruction tuning is like ==training a golden retriever== with homework.\nYou say ==“summarize this,”==\nand it learns not to bring you a tennis ball.\n\nIt turns a base model into a better helper.\nYou meet it inside chatbots and AI assistants.", "relationsNarrative": "Pretraining\nAfter Pretraining, instruction tuning teaches the model to follow requests.\n\nFine-tuning\nInstruction tuning is a common kind of Fine-tuning.\n\nRLHF\nInstruction tuning teaches request-following before RLHF teaches preferences.\n\nBase model\nInstruction tuning turns a Base model into a better chat helper.", "relations": { "pretraining": { "label": "comes after …", "note": "Models usually pretrain first, then learn to follow instructions." }, "fine-tuning": { "label": "is a common kind of …", "note": "Instruction tuning is a fine-tuning method." }, "rlhf": { "label": "often comes before …", "note": "It teaches the model to follow requests before RLHF teaches preferences." }, "base-model": { "label": "turns … into assistants", "note": "It turns a general base model into a chat-ready helper." } } }, "zh": { "fullName": "指令微调", "factExplain": "用指令—回答数据把模型调成更会听话的训练方式。", "humanExplain": "指令微调像给实习生带教：让他按“做表、改稿、列提纲”反复练，慢慢不再一听需求就自由发挥。\n\n常把基础模型调成更会按要求回复的助手。", "humanExplainDisplay": "指令微调像给实习生==带教==：\n让他按“做表、改稿、列提纲”\n反复练，慢慢不再==一听需求就自由发挥==。\n\n常把基础模型调成\n更会按要求回复的助手。", "relationsNarrative": "Pretraining\n它通常接在预训练之后，给模型补上“听指令”这一步。\n\nFine-tuning\n它本质上属于微调，是微调里很常见的一种做法。\n\nRLHF\n它常作为 RLHF 的前一步，先让模型学会按要求回答。\n\nBase model\n它把偏通用的基础模型，调成更像聊天助手的样子。", "relations": { "pretraining": { "label": "接在…之后", "note": "通常先预训练，再学会按指令作答。" }, "fine-tuning": { "label": "属于…常见形式", "note": "它本质上就是一种微调方法。" }, "rlhf": { "label": "常在…之前", "note": "先学会听话，再进一步学偏好。" }, "base-model": { "label": "把…调成助手", "note": "把通用底模调成可对话模型。" } } } } }, { "id": "interactive-video-generation", "name": "Interactive Video Generation", "layer": "L4", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "ai-long-video-generation" }, { "to": "generative-games" }, { "to": "world-model" }, { "to": "diffusion" } ], "track": "today", "i18n": { "en": { "fullName": "Interactive Video Generation", "factExplain": "Video AI that changes the scene based on what the user does.", "humanExplain": "It is like a movie with a game controller. Press B, and the hero jumps into a fountain.\n\nYou meet it in games, ads, and lessons. It turns video into something you can play with.", "humanExplainDisplay": "It is like a ==movie with a game controller==.\nPress B,\nand the hero ==jumps into a fountain==.\n\nYou meet it in games, ads, and lessons.\nIt turns video into something you can play with.", "relationsNarrative": "AI Long-Video Generation\nAI Long-Video Generation aims for length, and Interactive Video Generation aims for response.\n\nGenerative Games\nInteractive Video Generation creates the moving images for Generative Games.\n\nWorld model\nA World model helps actions cause sensible next scenes.\n\nDiffusion\nDiffusion often turns control signals into video frames.", "relations": { "ai-long-video-generation": { "label": "makes … controllable", "note": "Long video focuses on length. This focuses on response." }, "generative-games": { "label": "generates visuals for …", "note": "Playable worlds need video made as you play." }, "world-model": { "label": "uses … to predict changes", "note": "The world model makes each action lead to a sensible next scene." }, "diffusion": { "label": "often uses … for frames", "note": "Diffusion often turns control signals into detailed video frames." } } }, "zh": { "fullName": "互动视频生成", "factExplain": "根据用户操作动态生成或编辑视频内容。", "humanExplain": "互动视频生成像剧本杀 DM 开天眼：你一拍桌，剧情立刻改道。\n\n用于游戏、广告、教学，把视频变成可操作体验。", "humanExplainDisplay": "互动视频生成像\n==剧本杀 DM 开天眼==：\n你一拍桌，\n==剧情立刻改道==。\n\n用于游戏、广告、教学，\n把视频变成可操作体验。", "relationsNarrative": "AI Long-Video Generation\n长视频生成追求时长，它追求随操作变化。\n\nGenerative Games\n互动视频生成是生成式游戏的画面底座。\n\nWorld Model\n世界模型让用户动作引发合理后续画面。\n\nDiffusion\n扩散常用来把控制信号变成视频帧。", "relations": { "ai-long-video-generation": { "label": "让…可操控", "note": "长视频重连贯，互动视频重响应。" }, "generative-games": { "label": "生成…的画面", "note": "可玩的世界需要边操作边生成。" }, "world-model": { "label": "借…预测变化", "note": "世界模型负责让动作有后果。" }, "diffusion": { "label": "常用…生成帧", "note": "扩散常负责画面细节生成。" } } } } }, { "id": "inverse-reinforcement-learning", "name": "IRL", "layer": "L2", "era": "1998", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "imitation-learning" }, { "to": "markov-decision-process" } ], "track": "history", "i18n": { "en": { "fullName": "Inverse Reinforcement Learning", "factExplain": "A method for guessing the hidden reward rule from expert actions.", "humanExplain": "IRL is like copying Grandma’s cookies with no recipe. You watch her hands and guess her secret scorecard.\n\nIt works backward from examples to learn preferences for robots, self-driving cars, and game AI.", "humanExplainDisplay": "IRL is like copying Grandma’s cookies\nwith ==no recipe==.\nYou watch her hands\nand guess her ==secret scorecard==.\n\nIt works backward from examples\nto learn preferences for robots,\nself-driving cars, and game AI.", "relationsNarrative": "RL\nIRL learns the reward first, then RL finds the best plan.\n\nImitation Learning\nIRL also watches expert demos, but it cares about the hidden goal.\n\nMDP\nIRL often uses an MDP to define states, actions, and rewards.", "relations": { "reinforcement-learning": { "label": "learns the reward before …", "note": "It learns the reward first, then RL finds the best plan." }, "imitation-learning": { "label": "uses demos like …", "note": "It watches expert examples and guesses the goal behind them." }, "markov-decision-process": { "label": "models with …", "note": "An MDP often describes the states, actions, and rewards." } } }, "zh": { "fullName": "逆强化学习（Inverse Reinforcement Learning）", "factExplain": "从专家行为反推出奖励函数的学习方法。", "humanExplain": "IRL像偷学老师傅炒菜：不看菜谱，只看火候手法猜他心里那杆秤。\n\n从示范反推偏好，用于机器人、自动驾驶和游戏AI。", "humanExplainDisplay": "IRL像偷学老师傅炒菜：\n==不看菜谱==，\n只看火候手法\n猜他==心里那杆秤==。\n\n从示范反推偏好，\n用于机器人、自动驾驶\n和游戏AI。", "relationsNarrative": "Reinforcement Learning\n它反过来先学奖励函数，再用 RL 找策略。\n\nImitation Learning\n它同样看专家示范，但更关心背后的目标。\n\nMDP\n它通常在 MDP 框架里定义状态、动作和奖励。", "relations": { "reinforcement-learning": { "label": "反推…的奖励", "note": "先学奖励，再让智能体优化策略。" }, "imitation-learning": { "label": "借用…的示范", "note": "它从专家示范中推断目标。" }, "markov-decision-process": { "label": "依托…建模", "note": "状态、动作、奖励常由 MDP 表达。" } } } } }, { "id": "jailbreak", "name": "Jailbreak", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "prompt-injection" }, { "to": "alignment" }, { "to": "system-prompt" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "Jailbreak Prompt Attack", "factExplain": "A prompt attack that tries to make an AI ignore its safety rules.", "humanExplain": "A jailbreak is like telling the lunch lady, \"I just need a spoon.\" Really, you want to sneak behind the counter and ignore the rules.\n\nPeople use it to ask for banned content or unsafe steps. It is a common chatbot security risk.", "humanExplainDisplay": "A jailbreak is like telling the lunch lady,\n\"I just need a spoon.\"\nReally,\nyou want to ==sneak behind the counter==\nand ==ignore the rules==.\n\nPeople use it to ask for banned content\nor unsafe steps.\nIt is a common chatbot security risk.", "relationsNarrative": "Prompt injection\nJailbreaks are like prompt injection because both use input to steer the model off track.\n\nAlignment\nA jailbreak tries to get around the model's safety alignment.\n\nSystem prompt\nMany jailbreaks try to override or exploit the system prompt.\n\nAgent Security\nWhen an AI can use tools, a jailbreak becomes a more real security risk.", "relations": { "prompt-injection": { "label": "is a close cousin of …", "note": "Both use input to push the model off track." }, "alignment": { "label": "tries to bypass …", "note": "A jailbreak tests the model's safety boundaries." }, "system-prompt": { "label": "tries to override …", "note": "Many jailbreaks target the hidden instruction layer first." }, "agent-security": { "label": "is a key risk in …", "note": "The risk grows when the AI can use tools." } } }, "zh": { "fullName": "越狱提示攻击", "factExplain": "诱导模型绕过安全限制的提示攻击方式。", "humanExplain": "像跟小区门卫耍话术：嘴上说借个充电宝，实际是想混进楼里把整套规矩都绕开。\n\n常被用来套违规内容或危险指令，是聊天机器人防护里的高频风险。", "humanExplainDisplay": "像跟小区门卫\n==耍话术==：\n嘴上说借个充电宝，\n实际想把==整套规矩==绕开。\n\n常被用来套违规内容\n或危险指令；\n是聊天机器人防护里的\n高频风险。", "relationsNarrative": "Prompt Injection\n它和 Prompt Injection 很像，都是靠输入把模型带偏。\n\nAlignment\nJailbreak 的目标，就是绕过模型原本的安全对齐。\n\nSystem Prompt\n很多越狱会先试图覆盖或钻空子利用系统提示。\n\nAgent Security\n当模型能调工具时，越狱会变成更现实的安全风险。", "relations": { "prompt-injection": { "label": "常被归为…近亲", "note": "两者都靠输入操纵模型行为。" }, "alignment": { "label": "专门绕过…", "note": "越狱本质上在试探对齐边界。" }, "system-prompt": { "label": "试图覆盖…", "note": "很多越狱先攻击隐藏指令层。" }, "agent-security": { "label": "属于…重点风险", "note": "一旦接工具，危害通常更大。" } } } } }, { "id": "jax", "name": "JAX", "layer": "L5", "sublayer": "architecture", "era": "2018", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "automatic-differentiation" }, { "to": "tpu" }, { "to": "pytorch" } ], "track": "history", "i18n": { "en": { "fullName": "JAX", "factExplain": "A Python framework for automatic gradients and fast compiled math.", "humanExplain": "JAX is like a turbo button for math class. Write the formula, and it finds the slope, then races it on a GPU.\n\nResearchers use it to train models and run big tests. It helps Python feel much faster.", "humanExplainDisplay": "JAX is like a ==turbo button==\nfor math class.\nWrite the formula,\nand it ==finds the slope==,\nthen races it on a GPU.\n\nResearchers use it to train models\nand run big tests.\nIt helps Python feel much faster.", "relationsNarrative": "Autodiff\nAutodiff is a key JAX tool for training models.\n\nTPU\nJAX often compiles math so it runs fast on a TPU.\n\nPyTorch\nJAX and PyTorch are both common deep learning frameworks.", "relations": { "automatic-differentiation": { "label": "gets gradients with …", "note": "Autodiff turns formulas into gradients for training." }, "tpu": { "label": "runs computation on …", "note": "JAX often uses TPUs for fast parallel math." }, "pytorch": { "label": "compares with …", "note": "Both are popular frameworks for training deep learning models." } } }, "zh": { "fullName": "自动微分与加速计算框架", "factExplain": "一个支持自动微分和加速编译的 Python 计算框架。", "humanExplain": "JAX 像给公式叫网约车：你写数学，它自动求导，还直奔显卡高架。\n\n用于科研训练和大实验，让 Python 跑得快。", "humanExplainDisplay": "JAX 像==给公式叫网约车==：\n你写数学，\n它自动求导，\n还直奔==显卡高架==。\n\n用于科研训练和大实验，\n让 Python 跑得快。", "relationsNarrative": "Autodiff\n自动求导是 JAX 训练模型的关键能力。\n\nTPU\nJAX 常把计算编译后高效跑在 TPU 上。\n\nPyTorch\nJAX 和 PyTorch 都是常用深度学习框架。", "relations": { "automatic-differentiation": { "label": "用…求梯度", "note": "自动求导让公式直接变成梯度。" }, "tpu": { "label": "把计算跑上…", "note": "它常被用来发挥 TPU 的并行能力。" }, "pytorch": { "label": "对标…开发体验", "note": "两者都是研究者常用训练框架。" } } } } }, { "id": "k-means-clustering", "name": "K-Means Clustering", "layer": "L2", "era": "1957", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "clustering" }, { "to": "unsupervised-learning" }, { "to": "gaussian-mixture-model" }, { "to": "expectation-maximization" } ], "track": "history", "i18n": { "en": { "fullName": "k-Means Clustering", "factExplain": "An algorithm that groups data into k clusters by distance.", "humanExplain": "k-Means is like dropping k pizza boxes at a party. Everyone stands by the closest box. Then each box moves to the middle of its hungry crowd.\n\nYou meet it in customer groups and image compression. Pick k badly, and the groups look weird.", "humanExplainDisplay": "k-Means is like dropping ==k pizza boxes== at a party.\nEveryone stands by the ==closest box==.\nThen each box moves to the middle\nof its hungry crowd.\n\nYou meet it in customer groups and image compression.\nPick k badly,\nand the groups look weird.", "relationsNarrative": "Clustering\nk-Means is a classic way to group data automatically.\n\nUnsupervised Learning\nIt needs no human labels and groups by similarity.\n\nGMM\nGMM gives class probabilities, but k-Means gives one hard group.\n\nEM\nIt repeats assign then update until the centers stop moving.", "relations": { "clustering": { "label": "does …", "note": "It is one of the classic clustering algorithms." }, "unsupervised-learning": { "label": "belongs to …", "note": "It needs no labels and groups data by similarity." }, "gaussian-mixture-model": { "label": "contrasts with …", "note": "GMM gives soft probabilities; k-Means gives hard groups." }, "expectation-maximization": { "label": "works like …", "note": "It repeats assigning points and updating centers." } } }, "zh": { "fullName": "k均值聚类", "factExplain": "一种按距离把数据分成 k 类的聚类算法。", "humanExplain": "k均值像广场舞站队：先插几面旗，谁离得近跟谁，旗子再挪到队中央。\n\n用于用户分群、图像压缩；k选错就分得别扭。", "humanExplainDisplay": "k均值像\n==广场舞站队==：\n先插几面旗，谁离得近跟谁，\n==旗子再挪到队中央==。\n\n用于用户分群、图像压缩；\nk选错\n就分得别扭。", "relationsNarrative": "Clustering\n它是把数据自动分组的经典做法。\n\nUnsupervised Learning\n它不需要人工标签，只看数据相似度。\n\nGMM\nGMM 给每类概率，它只给一个归属。\n\nEM\n它也反复“分配—更新”，直到中心稳定。", "relations": { "clustering": { "label": "实现…", "note": "它是最经典的聚类算法之一。" }, "unsupervised-learning": { "label": "属于…", "note": "无需标签，也能按相似性分组。" }, "gaussian-mixture-model": { "label": "对比…", "note": "GMM 给软概率，它给硬分组。" }, "expectation-maximization": { "label": "类似…", "note": "它反复分配样本，再更新中心。" } } } } }, { "id": "k-nearest-neighbors", "name": "KNN", "layer": "L2", "era": "1951", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "classification" }, { "to": "regression" }, { "to": "curse-of-dimensionality" } ], "track": "history", "i18n": { "en": { "fullName": "k-Nearest Neighbors", "factExplain": "KNN predicts by letting the k closest examples vote or average.", "humanExplain": "KNN is like choosing pizza by asking the nearest kids in the cafeteria. If three nearby kids yell “pepperoni,” KNN orders pepperoni.\n\nIt uses nearby examples to sort things or guess numbers. It can spot oddballs too, but big data makes neighbor hunting slow.", "humanExplainDisplay": "KNN is like choosing pizza\nby asking the ==nearest kids== in the cafeteria.\nIf three nearby kids yell “pepperoni,”\nKNN ==orders pepperoni==.\n\nIt uses nearby examples\nto sort things or guess numbers.\nIt can spot oddballs too,\nbut big data makes neighbor hunting slow.", "relationsNarrative": "Supervised Learning\nKNN is a supervised learning method that uses labeled examples directly.\n\nClassification\nKNN lets the nearest neighbors vote for the class.\n\nRegression\nKNN averages nearby values to predict a number.\n\nCurse of Dimensionality\nWith too many features, distance gets blurry and KNN can miss.", "relations": { "supervised-learning": { "label": "is a kind of …", "note": "KNN uses labeled examples to make predictions." }, "classification": { "label": "often does …", "note": "KNN classifies by letting nearby examples vote." }, "regression": { "label": "can also do …", "note": "KNN predicts numbers by averaging nearby values." }, "curse-of-dimensionality": { "label": "struggles with …", "note": "Too many features can make distance less useful." } } }, "zh": { "fullName": "k-最近邻算法", "factExplain": "按最近的 k 个样本投票或平均来预测。", "humanExplain": "KNN像小区租房估价：不翻黄历，先看隔壁几套租多少，再跟着开价。\n\n用于分类、回归和异常检测；数据大时找邻居慢。", "humanExplainDisplay": "KNN像小区租房估价：\n不翻黄历，先看隔壁几套租多少，\n再==跟着开价==。\n\n用于分类、回归和异常检测；\n数据大时\n找邻居慢。", "relationsNarrative": "Supervised Learning\nKNN 是一种直接用标注样本做预测的监督学习方法。\n\nClassification\n分类时，它让最近邻居投票决定类别。\n\nRegression\n回归时，它把邻居的数值做平均。\n\nCurse Of Dimensionality\n高维下距离会失真，KNN 容易掉准。", "relations": { "supervised-learning": { "label": "属于…", "note": "KNN 直接用标注样本做预测。" }, "classification": { "label": "常用于…", "note": "KNN 常靠邻居投票完成分类。" }, "regression": { "label": "也可做…", "note": "把邻居数值平均就能回归。" }, "curse-of-dimensionality": { "label": "容易遭遇…", "note": "维度太高会让距离变得不可靠。" } } } } }, { "id": "kalman-filter", "name": "Kalman Filter", "layer": "L2", "era": "1960", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "belief-state" }, { "to": "hidden-markov-model" }, { "to": "robotaxi" } ], "track": "history", "i18n": { "en": { "fullName": "Kalman Filter", "factExplain": "A method that updates a changing estimate using prediction and new measurements.", "humanExplain": "A Kalman Filter is like following a shaky phone map. You guess your next step, then check the blue dot before you walk into a fountain.\n\nIt tracks a moving state, even when sensors are noisy. You meet it in location systems, sensor fusion, and self-driving cars.", "humanExplainDisplay": "A Kalman Filter is like following\na ==shaky phone map==.\nYou guess your next step,\nthen check the ==blue dot==\nbefore you walk into a fountain.\n\nIt tracks a moving state,\neven when sensors are noisy.\nYou meet it in location systems,\nsensor fusion,\nand self-driving cars.", "relationsNarrative": "Belief State\nA Kalman Filter updates the current belief with prediction and measurement.\n\nHMM\nBoth estimate hidden states, but a Kalman Filter handles continuous numbers.\n\nRobotaxi\nRobotaxis use Kalman Filters for localization and sensor fusion.", "relations": { "belief-state": { "label": "keeps updating …", "note": "It adds each new measurement to the current estimate." }, "hidden-markov-model": { "label": "also estimates hidden state like …", "note": "Both estimate states you cannot see directly." }, "robotaxi": { "label": "supports … localization", "note": "Self-driving cars use it to combine sensor data." } } }, "zh": { "fullName": "卡尔曼滤波", "factExplain": "一种用观测与预测递推估计状态的方法。", "humanExplain": "它像雨天骑车认路：前面靠感觉估方向，旁边再瞄路标修一把，才不至于越骑越偏。\n\n常用于定位、传感器融合和自动驾驶，能压住噪声、稳住连续状态估计。", "humanExplainDisplay": "它像雨天骑车认路：\n前面靠感觉估方向，\n旁边再瞄==路标==修一把，\n才不至于越骑越==偏==。\n\n常用于定位、传感器融合\n和自动驾驶，\n能压住噪声、稳住连续状态估计。", "relationsNarrative": "Belief State\n它用预测加观测，不断更新当前状态判断。\n\nHidden Markov Model\n两者都估计隐藏状态，但它更偏连续数值。\n\nRobotaxi\n自动驾驶常用它做定位与传感器融合。", "relations": { "belief-state": { "label": "持续更新…", "note": "它把新观测折进当前状态估计。" }, "hidden-markov-model": { "label": "同属时序估计", "note": "两者都处理看不全的隐藏状态。" }, "robotaxi": { "label": "支撑…定位", "note": "自动驾驶常靠它融合多路传感器。" } } } } }, { "id": "kernel-method", "name": "Kernel Method", "layer": "L2", "era": "1992", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "support-vector-machine" }, { "to": "classification" }, { "to": "regression" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Kernel Method", "factExplain": "A method that uses kernel functions to handle curved patterns in higher dimensions.", "humanExplain": "A kernel method is like a pop-up book for messy data. Flat blobs pop into shape, like a paper castle in math class.\n\nIt helps models handle curved patterns. You meet it in SVMs, classification, and regression, often with small data.", "humanExplainDisplay": "A kernel method is like a ==pop-up book== for messy data.\nFlat blobs ==pop into shape==,\nlike a paper castle in math class.\n\nIt helps models handle curved patterns.\nYou meet it in SVMs, classification, and regression,\noften with small data.", "relationsNarrative": "SVM\nSVM is one of the most classic uses of kernel methods.\n\nClassification\nKernel methods help with classes a straight line cannot split.\n\nRegression\nKernel tricks can predict numbers, not just classes.\n\nDeep Learning\nWith small data, people often compare kernel methods with Deep Learning.", "relations": { "support-vector-machine": { "label": "often used by …", "note": "SVM is the classic home for kernel methods." }, "classification": { "label": "often used for …", "note": "Kernel methods handle curved class borders well." }, "regression": { "label": "can also do …", "note": "Kernel tricks can help predict numbers too." }, "deep-learning": { "label": "often compared with …", "note": "On small data, people often compare it with deep learning." } } }, "zh": { "fullName": "核方法", "factExplain": "用核函数在高维空间处理非线性关系的方法。", "humanExplain": "核方法像武侠里借力打力，表面只出一掌，背后却把对手的弯招拐劲都顺手化开了。\n\n常用于分类、回归和小样本任务，小数据时往往很顺手。", "humanExplainDisplay": "核方法像武侠里==借力打力==，\n表面只出一掌，\n背后却把对手的==弯招拐劲==\n都顺手化开了。\n\n常用于分类、回归和小样本任务，\n小数据时往往很顺手。", "relationsNarrative": "Support-vector-machine\n支持向量机是核方法最经典、最常见的应用之一。\n\nClassification\n核方法常用来处理线性分不开的分类问题。\n\nRegression\n核技巧不只做分类，也能用于回归预测。\n\nDeep-learning\n在小样本场景里，它常被拿来和深度学习比较。", "relations": { "support-vector-machine": { "label": "常被…采用", "note": "支持向量机是核方法的经典代表。" }, "classification": { "label": "常用于…", "note": "核方法很适合处理非线性分类边界。" }, "regression": { "label": "也可做…", "note": "核技巧也能用于回归预测任务。" }, "deep-learning": { "label": "常与…对比", "note": "小数据场景下常拿它和深度学习比较。" } } } } }, { "id": "kimi-work", "name": "Kimi Work", "layer": "L5", "sublayer": "product", "era": "2026", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "copilot" }, { "to": "computer-use" }, { "to": "document-parsing" } ], "track": "history", "i18n": { "en": { "fullName": "Kimi Work", "factExplain": "An AI assistant product built to handle everyday office tasks.", "humanExplain": "Kimi Work is the office teammate with a desk-sized inbox tray. Toss it a messy spreadsheet, and it starts fixing the pile before lunch.\n\nPeople use it to sort files and draft reports. It also moves repeat office steps along.", "humanExplainDisplay": "Kimi Work is the office teammate\nwith a ==desk-sized inbox tray==.\nToss it a ==messy spreadsheet==,\nand it starts fixing the pile before lunch.\n\nPeople use it to sort files and draft reports.\nIt also moves repeat office steps along.", "relationsNarrative": "Agent\nKimi Work is an Agent product built for office work.\n\nCopilot\nKimi Work and Copilot both work beside people in daily workflows.\n\nComputer use\nOffice automation often needs Computer use to operate software for the user.\n\nDocument parsing\nKimi Work often needs Doc parsing before it can read PDFs, sheets, or contracts.", "relations": { "agent": { "label": "packages … for office work", "note": "It turns Agent skills into an office assistant product." }, "copilot": { "label": "works like …", "note": "Both work beside people inside daily work tasks." }, "computer-use": { "label": "may use … to act", "note": "Office tasks often need the AI to click and type in apps." }, "document-parsing": { "label": "uses … to read files", "note": "It must pull text from documents before it can work on them." } } }, "zh": { "fullName": "Kimi 办公智能体", "factExplain": "一种面向办公任务的 AI 助手产品形态。", "humanExplain": "像职场里的救火队员，谁把材料、表格、汇报甩过来，它都能先接住再往前推。\n\n常拿来整理资料、写汇报和跑流程，适合先自动化重复办公任务。", "humanExplainDisplay": "像职场里的==救火队员==，\n谁把材料、表格、汇报甩过来，\n它都能先==接住==再往前推。\n\n常拿来整理资料、\n写汇报和跑流程，\n适合先自动化重复办公任务。", "relationsNarrative": "Agent\n它是 Agent 在办公场景里的产品化落地。\n\nCopilot\n它和 Copilot 都强调在人类工作流中协作。\n\nComputer use\n做办公自动化时，常需要代用户操作软件界面。\n\nDocument parsing\n读懂 PDF、表格、合同，常要先做文档解析。", "relations": { "agent": { "label": "属于…产品化", "note": "它把代理能力包装成办公助手。" }, "copilot": { "label": "接近…形态", "note": "都主打在人类工作流里协作。" }, "computer-use": { "label": "可能用…执行", "note": "处理办公任务时常要代操作界面。" }, "document-parsing": { "label": "依赖…读材料", "note": "先把文档内容拆出来才好处理。" } } } } }, { "id": "knowledge-graph", "name": "Knowledge Graph", "layer": "L2", "era": "2012", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "rag" }, { "to": "graph-search" }, { "to": "vector-db" } ], "track": "history", "i18n": { "en": { "fullName": "Knowledge Graph", "factExplain": "A structured way to store facts as things and the links between them.", "humanExplain": "A knowledge graph is like a school gossip map on a whiteboard. A string from Emma to Pizza Palace says “works at.”\n\nYou meet it in search and Q&A. It also helps recommendations and company knowledge bases. It turns loose facts into a web of clear links.", "humanExplainDisplay": "A knowledge graph is like a ==school gossip map==\non a whiteboard.\nA string from Emma to Pizza Palace\nsays ==“works at.”==\n\nYou meet it in search and Q&A.\nIt also helps recommendations\nand company knowledge bases.\nIt turns loose facts\ninto a web of clear links.", "relationsNarrative": "KR\nA knowledge graph is a classic KR method for organizing facts with links.\n\nRAG\nA knowledge graph gives RAG clear clues about things and links.\n\nGraph Search\nGraph Search finds paths and related facts inside the graph.\n\nVector-db\nVector-db finds similar content, and the graph shows clear relationships.", "relations": { "knowledge-representation": { "label": "is a type of …", "note": "It is a classic structured way to represent knowledge." }, "rag": { "label": "can support …", "note": "It gives RAG clear facts and clear links." }, "graph-search": { "label": "uses … to explore", "note": "It helps find paths and related facts in the graph." }, "vector-db": { "label": "works well with …", "note": "A graph stores clear links. Vector-db finds similar meaning." } } }, "zh": { "fullName": "知识图谱", "factExplain": "用节点和关系组织知识的结构化表示。", "humanExplain": "知识图谱像相亲角那面关系墙：谁是谁同学、谁跟谁同事、谁又是哪家亲戚，全能顺藤摸瓜。\n\n常用于搜索问答、推荐和企业知识管理，把零散事实连成关系网。", "humanExplainDisplay": "知识图谱像相亲角那面\n==关系墙==：\n谁是谁同学、谁跟谁同事、\n谁又是哪家亲戚，\n全能==顺藤摸瓜==。\n\n常用于搜索问答、推荐\n和企业知识管理，\n把零散事实连成关系网。", "relationsNarrative": "Knowledge Representation\n知识图谱是知识表示里的经典做法，用关系把事实组织起来。\n\nRAG\n它可给检索增强提供更明确的实体与关系线索。\n\nGraph Search\n图搜索常用于在图谱里找路径、邻居和关联链路。\n\nVector DB\n向量库擅长找相似内容，图谱擅长表达明确关系。", "relations": { "knowledge-representation": { "label": "属于…的一种", "note": "它是结构化表达知识的经典形式。" }, "rag": { "label": "可作为…底座", "note": "可为检索增强提供结构化事实。" }, "graph-search": { "label": "依赖…遍历", "note": "找路径和关系时常要做图搜索。" }, "vector-db": { "label": "可与…互补", "note": "图管关系，向量库管语义相似。" } } } } }, { "id": "knowledge-representation", "name": "KR", "layer": "L2", "era": "1970", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "representation-learning" }, { "to": "embedding" }, { "to": "rag" }, { "to": "world-model" } ], "track": "history", "i18n": { "en": { "fullName": "Knowledge Representation", "factExplain": "A way to code knowledge into structures a machine can use.", "humanExplain": "Knowledge Representation is a seating chart for facts. It tells AI milk belongs in the fridge, not in the sock drawer.\n\nIt decides how machines store knowledge and use it. You meet it in knowledge bases, search tools, and assistants.", "humanExplainDisplay": "Knowledge Representation is a ==seating chart== for facts.\nIt tells AI milk belongs ==in the fridge==,\nnot in the sock drawer.\n\nIt decides how machines store knowledge and use it.\nYou meet it in knowledge bases,\nsearch tools,\nand assistants.", "relationsNarrative": "Representation Learning\nKnowledge Representation sets the form. Representation Learning learns it.\n\nEmbedding\nAn Embedding is a common number form for Knowledge Representation.\n\nRAG\nRAG finds knowledge first. Clear representation helps it search well.\n\nWorld model\nA World model needs clear objects and links. Knowledge Representation helps with that.", "relations": { "representation-learning": { "label": "is what … learns", "note": "Knowledge Representation sets the form. Representation Learning learns that form." }, "embedding": { "label": "often becomes … vectors", "note": "Knowledge can be coded as vectors a computer can compare." }, "rag": { "label": "helps … fetch knowledge", "note": "Clear representation helps RAG find better facts before it answers." }, "world-model": { "label": "can ground a …", "note": "A world model needs organized objects and links." } } }, "zh": { "fullName": "Knowledge Representation／知识表示", "factExplain": "把知识编码成机器可处理结构的方法。", "humanExplain": "像图书馆上架，不只记得有这本书，还得按类目归架、相关的摆一起，不然书在馆里也等于丢了。\n\n它决定机器怎么存知识、做检索和推理，常见于知识库、搜索和助手。", "humanExplainDisplay": "像==图书馆上架==，\n不只记得有这本书，\n还得按类目归架、==相关的摆一起==，\n不然书在馆里也等于丢了。\n\n它决定机器怎么存知识、\n做检索和推理，\n常见于知识库、搜索和助手。", "relationsNarrative": "Representation-learning\n知识表示关注“怎么表达知识”，表示学习关注“怎么学出这种表达”。\n\nEmbedding\n嵌入是知识表示的一种常见数值化形式。\n\nRAG\nRAG 先找到知识再作答，表示方式会影响检索效果。\n\nWorld model\n世界模型要理解环境，离不开对对象和关系的表示。", "relations": { "representation-learning": { "label": "是…的表达对象", "note": "前者讲知识内容，后者讲如何学表示。" }, "embedding": { "label": "常落成…向量", "note": "知识可被编码成便于计算的向量。" }, "rag": { "label": "支撑…取知识", "note": "表示清楚，检索和生成才更靠谱。" }, "world-model": { "label": "可构成…底座", "note": "世界模型需要组织对象与关系。" } } } } }, { "id": "kullback-leibler-divergence", "name": "KL Divergence", "layer": "L2", "era": "1951", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "cross-entropy-loss" }, { "to": "rlhf" }, { "to": "distillation" } ], "track": "history", "i18n": { "en": { "fullName": "Kullback-Leibler Divergence", "factExplain": "A one-way score for how different two probability distributions are.", "humanExplain": "You were 99% sure the vending machine had cookies. It drops kale chips. KL Divergence is the betrayal score.\n\nIt compares one set of chances with another. People use it to keep a model from drifting too far.", "humanExplainDisplay": "You were ==99% sure==\nthe vending machine had cookies.\nIt drops ==kale chips==.\nKL Divergence is the betrayal score.\n\nIt compares one set of chances\nwith another.\nPeople use it to keep a model\nfrom drifting too far.", "relationsNarrative": "Cross-entropy Loss\nCross-entropy is closely tied to KL divergence.\n\nRLHF\nRLHF often uses KL to keep policy updates from moving too far.\n\nDistillation\nDistillation uses KL to measure how close the student is to the teacher.", "relations": { "cross-entropy-loss": { "label": "helps make up …", "note": "Cross-entropy equals entropy plus KL divergence." }, "rlhf": { "label": "keeps … from drifting", "note": "RLHF uses KL to keep a policy near the reference model." }, "distillation": { "label": "measures … copying", "note": "Distillation uses KL to match the teacher's output probabilities." } } }, "zh": { "fullName": "Kullback-Leibler 散度／相对熵", "factExplain": "衡量两个概率分布差异的非对称指标。", "humanExplain": "像网购前你认准这家“稳得很”，结果快递到手货不对板；你原先越笃定，翻车时就越扎心。\n\n常用来约束模型别偏太远，也衡量分布差异。", "humanExplainDisplay": "像网购前你认准\n这家“==稳得很==”，\n结果快递到手==货不对板==；\n你原先越笃定，\n翻车时就越扎心。\n\n常用来约束模型别偏太远，\n也衡量分布差异。", "relationsNarrative": "Cross-entropy-loss\n交叉熵与它密切相关，常差一个目标分布的熵。\n\nRLHF\n在 RLHF 里，它常用于限制策略更新别偏太远。\n\nDistillation\n知识蒸馏常用它衡量学生对老师分布的贴近程度。", "relations": { "cross-entropy-loss": { "label": "构成…的一部分", "note": "交叉熵可看作熵加上它。" }, "rlhf": { "label": "约束…别跑偏", "note": "常拿它限制策略别偏离参考模型。" }, "distillation": { "label": "衡量…模仿程度", "note": "学生模型常用它追老师分布。" } } } } }, { "id": "kv-cache", "name": "KV cache", "layer": "L5", "sublayer": "compute", "era": "2017", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "attention" }, { "to": "inference" }, { "to": "vram" }, { "to": "continuous-batching" } ], "track": "today", "i18n": { "en": { "fullName": "Key-Value Cache", "factExplain": "A cache for Attention key-value data from already generated tokens.", "humanExplain": "KV cache is like a pizza shop leaving your order slip on the counter. When you add garlic bread, they do not ask for your whole life story again.\n\nIt speeds up long text generation. You meet it in chatbots and local LLMs.", "humanExplainDisplay": "KV cache is like a pizza shop\nleaving your ==order slip== on the counter.\nWhen you add garlic bread,\nthey do not ask for your ==whole life story== again.\n\nIt speeds up long text generation.\nYou meet it in chatbots\nand local LLMs.", "relationsNarrative": "Attention\nKV cache stores past keys and values for Attention, so they are not recomputed.\n\nInference\nKV cache speeds up Inference during token-by-token generation.\n\nVRAM\nKV cache uses VRAM, and long context can use a lot.\n\nContinuous batching\nOnline inference often uses KV cache with Continuous batching.", "relations": { "attention": { "label": "caches … results", "note": "It stores old keys and values used by Attention." }, "inference": { "label": "speeds up …", "note": "It cuts repeated work during token-by-token generation." }, "vram": { "label": "uses … space", "note": "Longer context usually means more VRAM for the cache." }, "continuous-batching": { "label": "works with …", "note": "Online inference often pairs it with Continuous batching." } } }, "zh": { "fullName": "Key-Value 缓存", "factExplain": "保存已生成 token 的注意力中间结果的缓存。", "humanExplain": "它最像聊天时有个机灵秘书，前面说过的话先==记在小本上==，后面续聊就不用每次从头翻旧账。\n\n它能加快长文本生成，常用于聊天机器人和本地大模型推理。", "humanExplainDisplay": "它最像聊天时有个机灵秘书，\n前面说过的话先\n==记在小本上==，\n后面续聊就不用每次\n==从头翻旧账==。\n\n它能加快长文本生成，\n常用于聊天机器人\n和本地大模型推理。", "relationsNarrative": "Attention\nKV cache 保存 Attention 的历史键和值，避免重复重算。\n\nInference\nKV cache 主要用于推理生成阶段，加快逐 token 输出。\n\nVRAM\nKV cache 会占用显存，长上下文时开销尤其明显。\n\nContinuous batching\n在线推理常把 KV cache 与动态批处理一起使用。", "relations": { "attention": { "label": "缓存…结果", "note": "它保存 Attention 用过的历史键值。" }, "inference": { "label": "加速…过程", "note": "主要在生成阶段减少重复计算。" }, "vram": { "label": "占用…空间", "note": "上下文越长，显存压力通常越大。" }, "continuous-batching": { "label": "配合…调度", "note": "在线推理常结合它提升并发表现。" } } } } }, { "id": "language-modeling", "name": "LM", "layer": "L2", "era": "1980s–1990s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "token" }, { "to": "llm" }, { "to": "transformer" } ], "track": "today", "i18n": { "en": { "fullName": "Language Modeling", "factExplain": "A task where a model learns language patterns by predicting the next piece of text.", "humanExplain": "It is like a never-ending fill-in-the-blank quiz. The AI sees some words. Then it guesses the next one.\n\nThis builds the base skill for writing text. You see it in pretraining, chat, and Q&A.", "humanExplainDisplay": "It is like a ==never-ending fill-in-the-blank quiz==.\nThe AI sees some words.\nThen it ==guesses the next one==.\n\nThis builds the base skill for writing text.\nYou see it in pretraining, chat, and Q&A.", "relationsNarrative": "Pretraining\nLanguage modeling is one of the main learning tasks in pretraining.\n\nToken\nLanguage modeling usually predicts the next token from the text before it.\n\nLLM\nAn LLM’s writing, Q&A, and chat skills are built on language modeling.\n\nTransformer\nTransformer is the most common setup for modern language modeling.", "relations": { "pretraining": { "label": "sits at the core of …", "note": "Pretraining often uses language modeling to learn language patterns." }, "token": { "label": "predicts … one by one", "note": "Language modeling usually predicts the next token from earlier text." }, "llm": { "label": "builds the base of …", "note": "An LLM first gets its text-making skill from language modeling." }, "transformer": { "label": "is often built with …", "note": "Modern language modeling is usually done with a Transformer." } } }, "zh": { "fullName": "Language modeling｜语言建模", "factExplain": "让模型学习并预测语言序列规律的任务。", "humanExplain": "语言建模像输入法的老熟人，你刚打“今晚吃”，它已经开始猜烧烤还是火锅。\n\n它是预训练的基本练习，支撑补全、翻译和聊天。", "humanExplainDisplay": "语言建模像==输入法的老熟人==，\n你刚打“今晚吃”，\n它已经开始==猜烧烤还是火锅==。\n\n它是预训练的基本练习，\n支撑补全、翻译和聊天。", "relationsNarrative": "Pretraining\n语言建模是预训练阶段最核心的学习任务之一。\n\nToken\n语言建模通常根据前文逐个预测下一个 Token。\n\nLLM\nLLM 的续写、问答和聊天能力建立在语言建模上。\n\nTransformer\nTransformer 是现代语言建模最常见的实现架构。", "relations": { "pretraining": { "label": "构成…核心任务", "note": "预训练常靠语言建模来学会语言规律。" }, "token": { "label": "逐个预测…", "note": "语言建模通常以前文预测下一个 token。" }, "llm": { "label": "奠定…能力底座", "note": "LLM 的生成能力首先来自语言建模。" }, "transformer": { "label": "常由…来实现", "note": "现代语言建模多用 Transformer 架构。" } } } } }, { "id": "lasso", "name": "Lasso", "layer": "L2", "era": "1996", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "regression" }, { "to": "parameter" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Least Absolute Shrinkage and Selection Operator", "factExplain": "A linear regression method that uses L1 regularization to set some weights to zero.", "humanExplain": "Lasso is a strict closet clean-out. If a shirt never gets worn, it gets booted from the drawer.\n\nIt keeps strong inputs and pushes weak ones to zero. You meet it in messy spreadsheets with many columns and not many rows.", "humanExplainDisplay": "Lasso is a ==strict closet clean-out==.\nIf a shirt never gets worn,\nit gets ==booted from the drawer==.\n\nIt keeps strong inputs\nand pushes weak ones to zero.\nYou meet it in messy spreadsheets\nwith many columns and not many rows.", "relationsNarrative": "Regularization\nLasso is a regularization method that uses a constraint to avoid overfitting.\n\nRegression\nLasso is often used for regression, especially to pick useful variables.\n\nParameter\nLasso can push some parameters to zero, so weak inputs get removed.\n\nBias-Variance Tradeoff\nLasso adds bias to reduce variance.", "relations": { "regularization": { "label": "belongs to …", "note": "Lasso is a classic regularization method." }, "regression": { "label": "is used for …", "note": "Lasso is a constrained version of regression." }, "parameter": { "label": "pushes … to zero", "note": "Lasso can make some parameters exactly zero." }, "bias-variance-tradeoff": { "label": "changes the …", "note": "Lasso adds bias to lower variance and behave more steadily." } } }, "zh": { "fullName": "最小绝对收缩与选择算子", "factExplain": "一种带 L1 正则的线性回归方法。", "humanExplain": "Lasso像收拾行李赶高铁，没用的统统别塞箱子，最后留下那几样，才是真要带上路的。\n\n常用于特征筛选和高维回归，变量多样本少时尤其好用。", "humanExplainDisplay": "Lasso像收拾行李赶高铁，\n没用的统统==别塞箱子==，\n最后留下那几样，\n才是真要==带上路的==。\n\n常用于特征筛选和高维回归，\n变量多样本少时尤其好用。", "relationsNarrative": "Regularization\nLasso 是正则化方法，用约束防止模型过度贴数据。\n\nRegression\n它常用于回归任务，尤其适合做变量筛选。\n\nParameter\nLasso 会把部分参数压成零，自动删掉弱特征。\n\nBias-Variance Tradeoff\n它通过增加偏差，来换取更低方差。", "relations": { "regularization": { "label": "属于…方法", "note": "它是经典正则化做法之一。" }, "regression": { "label": "常用于…任务", "note": "它本质上是回归模型的约束版。" }, "parameter": { "label": "把…压到零", "note": "会让部分参数直接变成零。" }, "bias-variance-tradeoff": { "label": "影响…平衡", "note": "靠加约束换更稳的泛化表现。" } } } } }, { "id": "latent-diffusion-model", "name": "LDM", "layer": "L3", "era": "2021", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "variational-autoencoder" }, { "to": "text-to-image-generation" }, { "to": "u-net" } ], "track": "history", "i18n": { "en": { "fullName": "Latent Diffusion Model", "factExplain": "A model that creates images by diffusing in latent space.", "humanExplain": "LDM is like drawing on a compressed draft. It denoises the picture inside a small latent space first, then decodes it back into the full image.\n\nIt powers many sharp text-to-image tools. It runs faster and uses less graphics memory.", "humanExplainDisplay": "LDM is like drawing on a ==compressed draft==.\nIt denoises inside a small ==latent space== first,\nthen decodes it back into the full image.\n\nIt powers many sharp text-to-image tools.\nIt runs faster and uses less graphics memory.", "relationsNarrative": "Diffusion\nLDM moves the diffusion process from pixels into latent space.\n\nVAE\nA VAE often compresses an image into latent space and decodes it back.\n\nText-to-Image Generation\nLDM is the speed base for many high-res text-to-image models.\n\nU-Net\nA U-Net often predicts noise in latent space and removes it step by step.", "relations": { "diffusion": { "label": "moves … into latent space", "note": "LDM does diffusion inside a compressed image form." }, "variational-autoencoder": { "label": "compresses images with …", "note": "A VAE often moves images into and out of latent space." }, "text-to-image-generation": { "label": "powers …", "note": "Many text-to-image systems use LDM to run faster." }, "u-net": { "label": "predicts noise with …", "note": "A U-Net often removes noise step by step in latent space." } } }, "zh": { "fullName": "潜空间扩散模型", "factExplain": "在潜空间中执行扩散生成的模型。", "humanExplain": "LDM 像在压缩草稿上作画：先在小小的潜空间里去噪成形，最后一步解码还原成完整图像。\n\n支撑高清文生图，让生成更快、更省显存。", "humanExplainDisplay": "LDM 像在\n==压缩草稿==上作画：\n先在潜空间里去噪成形，\n最后一步==解码还原==成完整图像。\n\n支撑高清文生图，\n让生成更快、\n更省显存。", "relationsNarrative": "Diffusion\nLDM 把扩散过程从像素空间搬到潜空间。\n\nVAE\nVAE 常负责把图像压缩进潜空间再解码。\n\nText-to-Image Generation\nLDM 是许多高清文生图模型的效率底座。\n\nU-Net\nU-Net 常在潜空间里预测噪声并逐步去噪。", "relations": { "diffusion": { "label": "把…搬进潜空间", "note": "LDM 在压缩表示里做扩散。" }, "variational-autoencoder": { "label": "借…压缩图像", "note": "VAE 常负责潜空间互转。" }, "text-to-image-generation": { "label": "支撑…", "note": "许多文生图系统靠它提速。" }, "u-net": { "label": "用…预测噪声", "note": "U-Net 常在潜空间里去噪。" } } } } }, { "id": "latent-dirichlet-allocation", "name": "LDA", "layer": "L3", "era": "2003", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "latent-variable-model" }, { "to": "probabilistic-graphical-model" }, { "to": "unsupervised-learning" }, { "to": "bag-of-words" } ], "track": "history", "i18n": { "en": { "fullName": "Latent Dirichlet Allocation", "factExplain": "A probability model for finding hidden topics from document word counts.", "humanExplain": "LDA is a cafeteria trash detective. It sees wrappers, then guesses pizza day or taco day.\n\nIt helps group articles and scan online chatter. First, it gives each text a topic color.", "humanExplainDisplay": "LDA is a ==cafeteria trash detective==.\nIt sees ==wrappers==,\nthen guesses pizza day or taco day.\n\nIt helps group articles and scan online chatter.\nFirst, it gives each text a topic color.", "relationsNarrative": "Latent Model\nLDA treats each topic as a hidden variable.\n\nPGM\nLDA can be drawn as a probability graph of documents, topics, and words.\n\nUnsupervised Learning\nLDA finds topics in text without human labels.\n\nBag-of-Words\nLDA usually counts words first with Bag-of-Words.", "relations": { "latent-variable-model": { "label": "is a kind of …", "note": "The topics are hidden variables inside the model." }, "probabilistic-graphical-model": { "label": "can be drawn as …", "note": "It links documents, topics, and words with probabilities." }, "unsupervised-learning": { "label": "learns without …", "note": "It can find topics without human labels." }, "bag-of-words": { "label": "often starts with …", "note": "It usually turns each document into a bag of word counts." } } }, "zh": { "fullName": "Latent Dirichlet Allocation，潜在狄利克雷分配", "factExplain": "从文档词频中发现潜在主题的概率模型。", "humanExplain": "LDA像中医闻药渣：不问病历，凭药味估这方子几分清热、几分补气。\n\n用于文章聚类、舆情分析，先给文本打主题底色。", "humanExplainDisplay": "LDA像中医闻药渣：\n不问病历，\n凭药味估这方子\n==几分清热==、==几分补气==。\n\n用于文章聚类、舆情分析，\n先给文本\n打主题底色。", "relationsNarrative": "Latent Model\nLDA 把“主题”当作看不见的潜变量。\n\nPGM\nLDA 可表示成文档、主题、词的概率图。\n\nUnsupervised Learning\nLDA 不靠人工标签，从文本里自己找主题。\n\nBag-of-Words\nLDA 通常把文档先看成词袋来统计。", "relations": { "latent-variable-model": { "label": "属于…", "note": "主题是模型里的隐藏变量。" }, "probabilistic-graphical-model": { "label": "可表示为…", "note": "用概率关系连接文档、主题和词。" }, "unsupervised-learning": { "label": "依赖…", "note": "不用标签也能挖出主题。" }, "bag-of-words": { "label": "常基于…", "note": "先把文章化成词频袋子。" } } } } }, { "id": "latent-semantic-analysis", "name": "LSA", "layer": "L2", "era": "1990", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "matrix-factorization" }, { "to": "dimensionality-reduction" }, { "to": "information-retrieval" }, { "to": "distributional-semantics" } ], "track": "history", "seo": { "zh": { "title": "潜在语义分析 是什么?听聊天的红娘,一文看懂 — AI Rookies", "description": "用矩阵分解发现词与文档潜在语义的方法。人话版一分钟讲明白。" }, "en": { "title": "What Is LSA? The Gossip Librarian Shelves Us", "description": "A method that uses matrix factorization to find hidden meaning in words and documents. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Latent Semantic Analysis", "factExplain": "A method that uses matrix factorization to find hidden meaning in words and documents.", "humanExplain": "LSA is like a school librarian with excellent gossip. You say “gym.” I say “workout.” She shelves us together.\n\nYou meet it in search and recommendations. It groups text by meaning, not just exact words.", "humanExplainDisplay": "LSA is like a ==school librarian==\nwith excellent gossip.\nYou say “gym.”\nI say “workout.”\nShe ==shelves us together==.\n\nYou meet it in search and recommendations.\nIt groups text by meaning,\nnot just exact words.", "relationsNarrative": "Matrix Factorization\nLSA uses matrix factorization to shrink word-document links into a meaning space.\n\nDim. Reduction\nLSA is an early dim. reduction method for text meaning.\n\nIR\nLSA helped match documents by meaning, not just exact words.\n\nDist. Semantics\nLSA uses word co-occurrence to guess hidden meaning.", "relations": { "matrix-factorization": { "label": "compresses with …", "note": "It splits the word-document table into a smaller meaning space." }, "dimensionality-reduction": { "label": "is a … method", "note": "It keeps the main meaning with fewer dimensions." }, "information-retrieval": { "label": "improves … matching", "note": "Search can match meaning, not just shared keywords." }, "distributional-semantics": { "label": "uses the … idea", "note": "Words get meaning from the company they keep." } } }, "zh": { "fullName": "潜在语义分析", "factExplain": "用矩阵分解发现词与文档潜在语义的方法。", "humanExplain": "LSA像相亲红娘听聊天：一个说撸铁，一个说健身房，也知道都在聊运动。\n\n用于搜索、聚类和推荐，按意思而不是字面找内容。", "humanExplainDisplay": "LSA像相亲红娘听聊天：\n一个说撸铁，\n一个说健身房，\n也知道==都在聊运动==。\n\n用于搜索、聚类和推荐，\n按意思而不是字面找内容。", "relationsNarrative": "Matrix Factorization\nLSA 通过矩阵分解把词文关系压到低维空间。\n\nDimensionality Reduction\nLSA 是文本语义空间的早期降维方法。\n\nInformation Retrieval\nLSA 曾用于按语义而非字面匹配文档。\n\nDistributional Semantics\nLSA 用共现分布估计词语的潜在意义。", "relations": { "matrix-factorization": { "label": "用…压缩矩阵", "note": "把词-文档表拆成低维语义空间。" }, "dimensionality-reduction": { "label": "属于…方法", "note": "用少数维度保留主要语义结构。" }, "information-retrieval": { "label": "改进…匹配", "note": "让搜索不只盯着关键词重合。" }, "distributional-semantics": { "label": "实践…思想", "note": "词的意义来自上下文共现。" } } } } }, { "id": "latent-variable-model", "name": "Latent Model", "layer": "L3", "era": "1950s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "variational-autoencoder" }, { "to": "expectation-maximization" }, { "to": "gaussian-mixture-model" } ], "track": "history", "i18n": { "en": { "fullName": "Latent Variable Model", "factExplain": "A model that explains seen data with hidden causes.", "humanExplain": "It is like finding socks on the floor and crumbs on the couch. You never saw the dog party, but the clues are loud.\n\nIt helps AI group data and shrink data. It also helps AI make new examples from hidden patterns.", "humanExplainDisplay": "It is like finding ==socks on the floor==\nand ==crumbs on the couch==.\nYou never saw the dog party,\nbut the clues are loud.\n\nIt helps AI group data and shrink data.\nIt also helps AI make new examples\nfrom hidden patterns.", "relationsNarrative": "Generative Model\nA latent variable model often uses hidden variables to generate and explain seen data.\n\nVAE\nA VAE joins latent variable modeling with neural network training.\n\nEM\nEM often trains these models by guessing hidden variables and model settings in turns.\n\nGMM\nA GMM is a classic latent variable model with hidden group labels.", "relations": { "generative-model": { "label": "is a kind of …", "note": "Many generative models use hidden variables to organize data." }, "variational-autoencoder": { "label": "is built as …", "note": "A VAE is a neural-network version of a latent variable model." }, "expectation-maximization": { "label": "is often trained with …", "note": "EM estimates hidden variables and model settings in turns." }, "gaussian-mixture-model": { "label": "has classic example …", "note": "A GMM uses a hidden group label to explain each sample." } } }, "zh": { "fullName": "潜变量模型", "factExplain": "用看不见的隐藏变量来解释观测数据的模型。", "humanExplain": "看朋友圈像看天气预报截图：表面都差不多，真正带节奏的，往往是没明说的那股心情。\n\n它常用于聚类、降维和生成建模，帮助模型找出隐藏结构。", "humanExplainDisplay": "看朋友圈像看天气预报截图：\n表面都差不多，\n真正==带节奏==的，\n往往是没明说的\n那股==心情==。\n\n它常用于聚类、降维\n和生成建模，\n帮助模型找出隐藏结构。", "relationsNarrative": "Generative Model\n潜变量模型常用隐藏变量来生成和解释观测数据。\n\nVariational Autoencoder\nVAE 把潜变量模型和神经网络训练结合起来。\n\nExpectation-Maximization\nEM 常用于这类模型中交替估计隐藏变量和参数。\n\nGaussian Mixture Model\nGMM 是经典潜变量模型，用隐藏类别解释样本。", "relations": { "generative-model": { "label": "属于…一类", "note": "很多生成模型靠隐藏变量组织数据。" }, "variational-autoencoder": { "label": "被…具体实现", "note": "VAE 是潜变量模型的神经网络版本。" }, "expectation-maximization": { "label": "常用…训练", "note": "EM 常用来估计隐藏变量与参数。" }, "gaussian-mixture-model": { "label": "典型例子是…", "note": "GMM 用类别潜变量解释数据来源。" } } } } }, { "id": "layer-normalization", "name": "Layer Normalization", "layer": "L2", "era": "2016", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "batch-normalization" }, { "to": "transformer" }, { "to": "optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Layer Normalization", "factExplain": "A method that standardizes the features inside one data example.", "humanExplain": "Layer Normalization is like seasoning each dish separately. You fix the salt in this one pan — the next order can be mild or spicy, it does not touch your work.\n\nTransformers use it to keep training steady. Deep networks learn better when their numbers stop wobbling.", "humanExplainDisplay": "Layer Normalization is like\n==seasoning each dish separately==.\nYou fix the salt in ==this one pan==.\nThe next order can be mild or spicy.\nIt does not touch your work.\n\nTransformers use it to keep training steady.\nDeep networks learn better\nwhen their numbers stop wobbling.", "relationsNarrative": "Batch Normalization\nBoth steady training, but Layer Normalization does not use batch stats.\n\nTransformer\nTransformers often use Layer Normalization to keep deep layers steady.\n\nOptimization\nLayer Normalization makes training smoother and less wobbly.", "relations": { "batch-normalization": { "label": "is often compared with …", "note": "Both normalize values, but Batch Normalization uses batch stats." }, "transformer": { "label": "is often used by …", "note": "Transformers use it to keep deep training steady." }, "optimization": { "label": "helps steady …", "note": "It reduces number swings during training." } } }, "zh": { "fullName": "层归一化", "factExplain": "按单个样本的特征维度做标准化的方法。", "humanExplain": "层归一化像做菜时每道菜单独调味：咸淡只盯眼前这盘，隔壁桌点的是麻辣还是清蒸都不影响你。\n\n常用在 Transformer 里稳住训练，让深层网络更好学。", "humanExplainDisplay": "层归一化像做菜时\n每道菜单独==调味==：\n咸淡只盯眼前这盘，\n隔壁桌点的是麻辣还是清蒸\n都不影响你==稳稳出锅==。\n\n常用在 Transformer 里稳住训练，\n让深层网络更好学。", "relationsNarrative": "Batch Normalization\n两者都在稳训练，但它不依赖 batch 统计量。\n\nTransformer\nTransformer 广泛使用它来稳定层间激活分布。\n\nOptimization\n它能让优化过程更平稳，减少训练发飘。", "relations": { "batch-normalization": { "label": "常与…对比", "note": "两者都做归一化，但统计方式不同。" }, "transformer": { "label": "常被…采用", "note": "Transformer 普遍用它稳定深层训练。" }, "optimization": { "label": "帮助…稳定", "note": "它能缓和训练中的数值波动。" } } } } }, { "id": "legal-ai", "name": "Legal AI", "layer": "L5", "sublayer": "product", "era": "2010s", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "rag" }, { "to": "hallucination" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "Legal AI", "factExplain": "AI used to search, review, and analyze legal text.", "humanExplain": "Legal AI is the law office intern with turbo coffee. It reads piles of papers fast. But the lawyer still holds the pen.\n\nIt helps find legal text. It can review contracts and draft short summaries. It saves time. It does not replace legal judgment.", "humanExplainDisplay": "Legal AI is the ==law office intern==\nwith ==turbo coffee==.\nIt reads piles of papers fast.\nBut the lawyer still holds the pen.\n\nIt helps find legal text.\nIt can review contracts\nand draft short summaries.\nIt saves time.\nIt does not replace legal judgment.", "relationsNarrative": "LLM\nLegal AI uses an LLM to read, summarize, and answer legal questions.\n\nRAG\nRAG helps it find support in laws, cases, and contract files.\n\nHallucination\nLegal work cannot afford Hallucination. A fake case may look real.\n\nData-privacy\nLegal AI must protect Data-privacy when it handles files and contracts.", "relations": { "llm": { "label": "understands text with …", "note": "The LLM reads, summarizes, and answers legal questions." }, "rag": { "label": "checks sources with …", "note": "RAG links answers back to laws, cases, or files." }, "hallucination": { "label": "must block …", "note": "A fake law or case can mislead people fast." }, "data-privacy": { "label": "must protect …", "note": "Case files and contracts often hold sensitive information." } } }, "zh": { "fullName": "法律人工智能", "factExplain": "用于法律检索、审查和分析的 AI 应用。", "humanExplain": "法律 AI 是律所里的扫地僧实习生：翻案卷飞快，拍板签字还得师傅来。\n\n用于检索、审合同和摘要，提效不替代判断。", "humanExplainDisplay": "法律 AI 是律所里的\n==扫地僧实习生==：\n翻案卷飞快，\n拍板签字还得==师傅来==。\n\n用于检索、审合同和摘要，\n提效不替代判断。", "relationsNarrative": "LLM\n法律 AI 借助 LLM 阅读、摘要和回答法律问题。\n\nRAG\nRAG 帮它从法规、判例和合同库里找依据。\n\nHallucination\n法律场景最怕幻觉，把不存在的判例说得像真的。\n\nData-privacy\n它处理合同和案卷时，必须保护敏感信息。", "relations": { "llm": { "label": "用…理解法律文本", "note": "LLM 提供阅读、摘要和问答能力。" }, "rag": { "label": "用…查法规案例", "note": "RAG 让答案可追到法条来源。" }, "hallucination": { "label": "必须防住…", "note": "编错法条或判例会直接误导人。" }, "data-privacy": { "label": "必须守住…", "note": "案卷和合同常含高度敏感信息。" } } } } }, { "id": "lenet-5", "name": "LeNet-5", "layer": "L3", "era": "1998", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "mnist" }, { "to": "alexnet" }, { "to": "backpropagation" } ], "track": "history", "i18n": { "en": { "fullName": "LeNet-5", "factExplain": "An early CNN built to read handwritten numbers.", "humanExplain": "LeNet-5 is the old-school bank teller of AI. Even with wobbly check numbers, it can spot 0 through 9.\n\nIt was used for checks and ZIP code digits. It is still a starter example for learning CNNs.", "humanExplainDisplay": "LeNet-5 is the ==old-school bank teller== of AI.\nEven with ==wobbly check numbers==,\nit can spot 0 through 9.\n\nIt was used for checks and ZIP code digits.\nIt is still a starter example for learning CNNs.", "relationsNarrative": "CNN\nLeNet-5 is a classic early CNN and showed the design could read images.\n\nMNIST\nMNIST made LeNet-5 a starter model for handwritten digits.\n\nAlexNet\nAlexNet followed the CNN path and scaled it up.\n\nBackpropagation\nLeNet-5 uses Backpropagation to learn its settings end to end.", "relations": { "cnn": { "label": "is a classic early …", "note": "LeNet-5 is a classic early CNN design." }, "mnist": { "label": "practices on …", "note": "MNIST is its famous handwritten digit practice set." }, "alexnet": { "label": "inspired later …", "note": "AlexNet took the CNN path and made it much bigger." }, "backpropagation": { "label": "learns through …", "note": "Backpropagation trains its parts end to end." } } }, "zh": { "fullName": "早期手写数字识别卷积网络", "factExplain": "一种用于手写数字识别的早期卷积神经网络。", "humanExplain": "LeNet-5 像银行柜台老师傅：歪七扭八的数字，也能认出 0 到 9。\n\n用于支票、邮编数字识别，也是 CNN 入门样板。", "humanExplainDisplay": "LeNet-5 像\n==银行柜台老师傅==：\n歪七扭八的数字，\n也能==认出 0 到 9==。\n\n用于支票、\n邮编数字识别，\n也是 CNN 入门样板。", "relationsNarrative": "CNN\n它是早期经典 CNN，验证了卷积架构的价值。\n\nMNIST\nMNIST 让它成为手写数字识别的入门样板。\n\nAlexNet\nAlexNet 延续 CNN 路线，并把规模做大。\n\nBackpropagation\n它依靠反向传播端到端学习参数。", "relations": { "cnn": { "label": "代表早期…", "note": "它是经典早期 CNN 架构。" }, "mnist": { "label": "常在…上识别数字", "note": "MNIST 是它的标志性练习场。" }, "alexnet": { "label": "启发后来的…", "note": "AlexNet 把 CNN 推向大规模视觉。" }, "backpropagation": { "label": "依靠…训练", "note": "端到端训练离不开反向传播。" } } } } }, { "id": "lime", "name": "LIME", "layer": "L2", "era": "2016", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "explainable-ai" }, { "to": "shap" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Local Interpretable Model-agnostic Explanations", "factExplain": "A method using a simple nearby model to explain one prediction.", "humanExplain": "LIME is like checking one weird pizza slice. You inspect that bite, not the whole restaurant.\n\nIt explains one black-box decision at a time. You see it in loan checks and medical AI.", "humanExplainDisplay": "LIME is like checking ==one weird pizza slice==.\nYou inspect ==that bite==,\nnot the whole restaurant.\n\nIt explains one black-box decision at a time.\nYou see it in loan checks and medical AI.", "relationsNarrative": "XAI\nLIME is a common local explanation method in XAI.\n\nSHAP\nBoth explain one prediction. SHAP focuses more on consistent scores.\n\nClassification\nLIME often explains why a classifier picked one class.", "relations": { "explainable-ai": { "label": "is part of …", "note": "LIME is a common tool for explaining black-box predictions." }, "shap": { "label": "is often compared with …", "note": "Both explain why one prediction happened." }, "classification": { "label": "explains … results", "note": "LIME often shows why a model picked one class." } } }, "zh": { "fullName": "局部可解释模型无关解释", "factExplain": "用简单局部模型解释单次预测的方法。", "humanExplain": "LIME 像法官宣判：别背整部法典，只说这案子哪几条证据定输赢。\n\n用于解释黑箱单次判断，常见于风控医疗。", "humanExplainDisplay": "LIME 像法官宣判：\n==别背整部法典==，\n只说这案子，\n==哪几条证据定输赢==。\n\n用于解释黑箱单次判断，\n常见于风控医疗。", "relationsNarrative": "XAI\nLIME 是 XAI 中常见的局部解释方法。\n\nSHAP\n两者都解释单次预测，SHAP 更重一致性。\n\nClassification\n它常解释分类模型为何判成某一类。", "relations": { "explainable-ai": { "label": "属于…方法", "note": "它是解释黑箱预测的常用工具。" }, "shap": { "label": "常与…对比", "note": "两者都解释单次预测依据。" }, "classification": { "label": "解释…结果", "note": "常说明分类为何判成某类。" } } } } }, { "id": "lingbot-depth", "name": "LingBot-Depth", "layer": "L3", "era": "2026", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "spatial-intelligence" }, { "to": "embodied-ai" }, { "to": "masked-language-modeling" }, { "to": "self-supervised-learning" } ], "track": "today", "seo": { "en": { "title": "What Is LingBot-Depth? Robot Depth Sense From Plain Photos", "description": "Ant's open-source spatial model lets robots judge distance from ordinary photos — even glass and mirrors that blind normal sensors. A plain-English explainer." }, "zh": { "title": "LingBot-Depth 是什么?专治透明反光的深度感知,一文看懂 — AI Rookies", "description": "玻璃、镜子让传感器抓瞎,它光看普通照片就能判远近。蚂蚁灵波开源的空间感知模型,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "LingBot-Depth", "factExplain": "An open-source spatial model from Ant's Robbyant that lets robots judge distance from plain photos.", "humanExplain": "Depth cameras go blind on glass and mirrors. LingBot-Depth is like a carpenter who just eyeballs how thick a board is—it reads a plain photo and guesses how far things are.\n\nRobots use it to grab clear or shiny objects. It fills in the depth the sensor missed.", "humanExplainDisplay": "Depth cameras go blind\non ==glass and mirrors==.\nLingBot-Depth is like a carpenter\nwho just ==eyeballs how thick== a board is.\n\nRobots use it to grab\nclear or shiny objects.\nIt fills in the depth\nthe sensor missed.", "relationsNarrative": "Spatial Intelligence\nLingBot-Depth brings Spatial Intelligence into robot vision.\n\nEmbodied AI\nIt gives embodied robots a clearer sense of distance.\n\nMLM\nIt borrows the masked-modeling idea and applies it to depth maps.\n\nSSL\nHiding depth and asking the model to fill it in is self-supervised learning.", "relations": { "spatial-intelligence": { "label": "grounds …", "note": "It brings spatial intelligence into robot depth vision." }, "embodied-ai": { "label": "powers …", "note": "It gives embodied robots a clearer sense of distance." }, "masked-language-modeling": { "label": "borrows from …", "note": "It reuses the masked-modeling idea on depth maps." }, "self-supervised-learning": { "label": "trains with …", "note": "Hiding depth and filling it back in is self-supervised." } } }, "zh": { "fullName": "灵波空间感知模型", "factExplain": "蚂蚁灵波开源的空间感知模型，让机器人只看普通照片也能判断物体远近。", "humanExplain": "LingBot-Depth 像老木匠瞄一眼估厚薄：玻璃、镜子这些让传感器抓瞎的东西，它光看照片就判出远近。\n\n用于机器人抓取，专治透明反光物件的深度难题。", "humanExplainDisplay": "LingBot-Depth 像\n==老木匠瞄一眼估厚薄==：\n玻璃、镜子这些让\n传感器抓瞎的东西，\n它==光看照片就判出远近==。\n\n用于机器人抓取，\n专治透明反光物件的深度难题。", "relationsNarrative": "Spatial Intelligence\n它是空间智能在机器人视觉上的具体落地。\n\nEmbodied AI\n它给具身机器人补上「看清远近」这一环。\n\nMLM\n它借鉴掩码建模，把「遮住再猜」用在深度图上。\n\nSSL\n遮住深度让模型自己补，属于自监督训练。", "relations": { "spatial-intelligence": { "label": "落地…", "note": "它把空间智能做进机器人的深度视觉里。" }, "embodied-ai": { "label": "支撑…", "note": "给具身机器人补上「看清远近」这一环。" }, "masked-language-modeling": { "label": "借鉴…", "note": "把「遮住猜」的掩码建模思路用到深度图上。" }, "self-supervised-learning": { "label": "靠…训练", "note": "遮住深度让模型自己补，是一种自监督。" } } } } }, { "id": "lingbot-vla", "name": "LingBot-VLA", "layer": "L3", "era": "2026", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "vision-language-action-model-vla" }, { "to": "embodied-ai" }, { "to": "lingbot-depth" }, { "to": "foundation-model" } ], "track": "today", "seo": { "en": { "title": "What Is LingBot-VLA? One Robot Brain for 20+ Robot Bodies", "description": "Ant's open-source vision-language-action model, trained on 60,000 hours of real robot data — one brain that adapts across robots from 17 makers. Explained in plain English." }, "zh": { "title": "LingBot-VLA 是什么?蚂蚁灵波的具身基座模型,一文看懂 — AI Rookies", "description": "6 万小时真机数据练出的通用机器人大脑,一个底座适配 17 家厂商 20 多种机身。它和 LingBot-Depth 怎么分工?人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "LingBot-VLA", "factExplain": "An open-source robot brain from Ant's Robbyant that sees, listens, and acts across 20+ robot bodies from 17 makers.", "humanExplain": "LingBot-VLA is like a great all-around athlete. Show it a new sport, and it picks up the moves after a few tries.\n\nIt trained on 60,000 hours of real robot practice. Swap the robot body, and the brain still works.", "humanExplainDisplay": "LingBot-VLA is like a great\n==all-around athlete==.\nShow it a new sport,\nand it picks up the moves\n==after a few tries==.\n\nIt trained on 60,000 hours\nof real robot practice.\nSwap the robot body,\nand the brain still works.", "relationsNarrative": "VLA\nLingBot-VLA is an open-source base model on the vision-language-action path.\n\nEmbodied AI\nIt gives embodied robots one shared brain.\n\nLingBot-Depth\nIts sibling LingBot-Depth judges distance; LingBot-VLA understands and acts.\n\nFoundation-model\nIt acts like a foundation model for robots: one base, many bodies.", "relations": { "vision-language-action-model-vla": { "label": "is a …", "note": "It is an open-source base model on the VLA path." }, "embodied-ai": { "label": "powers …", "note": "It gives embodied robots one shared brain." }, "lingbot-depth": { "label": "teams up with …", "note": "Its sibling judges distance; it understands and acts." }, "foundation-model": { "label": "acts as … for robots", "note": "One base fits robots from many makers." } } }, "zh": { "fullName": "灵波具身基座模型", "factExplain": "蚂蚁灵波开源的「视觉-语言-动作」基座模型，一个大脑适配 17 家厂商 20 多种机器人。", "humanExplain": "LingBot-VLA 像十项全能运动员：跑跳投掷样样练过，换个新项目，摸几把就能上场。\n\n6 万小时真机数据练出的通用大脑，换机身不用从头教。", "humanExplainDisplay": "LingBot-VLA 像\n==十项全能运动员==：\n跑跳投掷样样练过，\n换个新项目，\n==摸几把就能上场==。\n\n6 万小时真机数据\n练出的通用大脑，\n换机身不用从头教。", "relationsNarrative": "VLA\n它是「视觉-语言-动作」路线的开源基座实现。\n\nEmbodied AI\n它给具身机器人提供一个通用大脑。\n\nLingBot-Depth\n同门的 LingBot-Depth 管看清远近，它管看懂了动手。\n\nFoundation-model\n它像机器人界的基座模型，一个底座适配多家机身。", "relations": { "vision-language-action-model-vla": { "label": "属于…", "note": "它是「视觉-语言-动作」路线的开源基座。" }, "embodied-ai": { "label": "充当…大脑", "note": "给具身机器人提供一个通用「大脑」。" }, "lingbot-depth": { "label": "搭档…", "note": "同门的它管看清远近，自己管看懂了动手。" }, "foundation-model": { "label": "是机器人界的…", "note": "一个底座适配多家厂商的机器人。" } } } } }, { "id": "lisp", "name": "Lisp", "layer": "L5", "sublayer": "product", "era": "1958", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "dartmouth-workshop" }, { "to": "knowledge-representation" }, { "to": "logic-programming" } ], "track": "history", "i18n": { "en": { "fullName": "LISt Processor", "factExplain": "A programming language for symbols and self-calling code.", "humanExplain": "Lisp looks like a coder spilled parentheses into a cereal bowl. Early AI grabbed a spoon.\n\nIt handles symbol rules and self-calling code. It powered many early AI systems.", "humanExplainDisplay": "Lisp looks like a coder spilled ==parentheses==\ninto a ==cereal bowl==.\nEarly AI grabbed a spoon.\n\nIt handles symbol rules\nand self-calling code.\nIt powered many early AI systems.", "relationsNarrative": "Dartmouth Workshop\nLisp grew in the early AI boom after the Dartmouth Workshop.\n\nKR\nLisp is good at working with symbol structures for KR experiments.\n\nLogic\nLisp and Logic both belong to the symbolic AI programming tradition.", "relations": { "dartmouth-workshop": { "label": "rode the … wave", "note": "After Dartmouth, Lisp became a key early AI tool." }, "knowledge-representation": { "label": "helps express …", "note": "Lisp is handy for working with symbolic knowledge." }, "logic-programming": { "label": "shares symbol roots with …", "note": "Both like rules and symbolic reasoning." } } }, "zh": { "fullName": "LISt Processor，表处理语言", "factExplain": "一种面向符号处理和递归的编程语言。", "humanExplain": "Lisp 像程序员盘佛珠：一颗颗括号套下去，早期 AI 靠它念逻辑经。\n\n擅长符号、规则和递归，撑起早期 AI 系统。", "humanExplainDisplay": "Lisp 像程序员==盘佛珠==：\n一颗颗括号套下去，\n早期 AI 靠它\n==念逻辑经==。\n\n擅长符号、规则和递归，\n撑起早期 AI 系统。", "relationsNarrative": "Dartmouth Workshop\nLisp 诞生于达特茅斯会议后的早期 AI 热潮。\n\nKR\nLisp 擅长操作符号结构，常用于知识表示实验。\n\nLogic\nLisp 与逻辑编程同属符号 AI 的编程传统。", "relations": { "dartmouth-workshop": { "label": "承接…热潮", "note": "达特茅斯后，它成了早期 AI 工具。" }, "knowledge-representation": { "label": "服务…表达", "note": "符号知识很适合用它操作。" }, "logic-programming": { "label": "同属符号传统", "note": "二者都偏爱规则和符号推理。" } } } } }, { "id": "llama-cpp", "name": "Llama.cpp", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "local-llm" }, { "to": "gguf" }, { "to": "quantization" }, { "to": "ollama" } ], "track": "history", "i18n": { "en": { "fullName": "llama.cpp", "factExplain": "An open-source tool for running big AI models fast on local devices.", "humanExplain": "llama.cpp is like fitting a food truck into a lunchbox. The AI still cooks, but now it sits on your laptop.\n\nPeople use it for offline chat and private AI at work. It saves cloud money, but your device may sweat.", "humanExplainDisplay": "llama.cpp is like fitting a ==food truck==\ninto a ==lunchbox==.\nThe AI still cooks,\nbut now it sits on your laptop.\n\nPeople use it for offline chat\nand private AI at work.\nIt saves cloud money,\nbut your device may sweat.", "relationsNarrative": "Local-LLM\nllama.cpp is a core engine for many Local-LLM setups.\n\nGGUF\nllama.cpp often loads GGUF model files for local use.\n\nQuantization\nQuantization makes models smaller, so llama.cpp can fit them on normal devices.\n\nOllama\nOllama often uses llama.cpp as one of its model engines.", "relations": { "local-llm": { "label": "helps power …", "note": "Many local AI setups use llama.cpp to run the model." }, "gguf": { "label": "often loads …", "note": "llama.cpp often loads local model files in GGUF format." }, "quantization": { "label": "gets lighter with …", "note": "Quantization shrinks models so llama.cpp can run them on normal devices." }, "ollama": { "label": "often runs under …", "note": "Ollama often uses llama.cpp as one engine for running models." } } }, "zh": { "fullName": "本地大语言模型推理框架", "factExplain": "一个让大模型在本地设备高效运行的开源推理项目。", "humanExplain": "本来得包场机房的大模型，被 llama.cpp 练成地铁通勤党：塞进笔记本，也能低调开工。\n\n常用来本地跑模型、离线聊天和私有部署，省云端成本，但更吃设备性能。", "humanExplainDisplay": "本来得包场机房的，\n被 llama.cpp 练成\n==地铁通勤党==：\n==塞进笔记本==也能开工。\n\n常用来本地跑模型、\n离线聊天和私有部署；\n省云端成本，\n但更吃设备性能。", "relationsNarrative": "Local-LLM\n它是很多本地大模型方案的核心运行底座。\n\nGGUF\n它常加载 GGUF 格式的模型文件来本地推理。\n\nQuantization\n量化能减小模型体积，帮它跑进普通设备。\n\nOllama\nOllama 常把它作为底层执行引擎之一。", "relations": { "local-llm": { "label": "支撑…落地", "note": "很多本地跑模型方案都靠它。" }, "gguf": { "label": "常配合…使用", "note": "它常加载这种本地模型文件格式。" }, "quantization": { "label": "依赖…减负", "note": "量化后模型更容易塞进普通设备。" }, "ollama": { "label": "常被…调用", "note": "不少本地工具底层会接它来跑模型。" } } } } }, { "id": "llm-as-a-judge", "name": "LLM-as-a-judge", "layer": "L4", "era": "2023", "publishedAt": "2026-05-29T16:08:01.207Z", "relations": [ { "to": "llm" }, { "to": "benchmark-contamination" }, { "to": "ai-bias" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "LLM-as-a-judge", "factExplain": "A way to use an LLM to score answers automatically.", "humanExplain": "It is like putting AI on a talent-show panel. It gives scores fast, but may still boo the best singer.\n\nPeople use it to grade model answers and tests. It saves time, but humans should check important scores.", "humanExplainDisplay": "It is like putting AI on a ==talent-show panel==.\nIt gives scores fast,\nbut may still ==boo the best singer==.\n\nPeople use it to grade model answers and tests.\nIt saves time,\nbut humans should check important scores.", "relationsNarrative": "LLM\nLLM-as-a-judge uses an LLM as an automatic scorer or comparer.\n\nBenchmark contamination\nBenchmark contamination can make LLM-as-a-judge scores look better than they are.\n\nAI-bias\nAI-bias in the judge model can make the scores unfair.\n\nHuman-in-the-loop\nHuman-in-the-loop helps catch bad scores in important cases.", "relations": { "llm": { "label": "uses … as judge", "note": "LLM-as-a-judge uses an LLM to score another output." }, "benchmark-contamination": { "label": "can be skewed by …", "note": "If the model has seen the test, the score may be false." }, "ai-bias": { "label": "can amplify …", "note": "A biased judge model can give biased scores." }, "human-in-the-loop": { "label": "often needs … review", "note": "Important scores still need a human safety check." } } }, "zh": { "fullName": "用大模型当评委", "factExplain": "用大模型自动评估回答质量的方法。", "humanExplain": "把 AI 请上评委席，打分比老师改卷还麻利；可它自己也会带主观，偶尔一本正经地判错。\n\n用于模型评测和答案打分，能省人力，关键结果仍宜人工复核。", "humanExplainDisplay": "把 AI 请上==评委席==，\n打分比老师改卷还麻利；\n可它自己也会带主观，\n偶尔一本正经地==判错==。\n\n用于模型评测和答案打分，\n能省人力，关键结果仍宜人工复核。", "relationsNarrative": "LLM\nLLM-as-a-judge 本质上是把 LLM 用作自动评分器或比较器。\n\nBenchmark contamination\n如果评测题目或答案被模型提前见过，LLM-as-a-judge 的分数可能失真。\n\nAI-bias\n当评委模型本身带有偏见时，评分结果也可能跟着偏。\n\nHuman-in-the-loop\n在高风险或重要评测中，通常要配合人工复核来兜底。", "relations": { "llm": { "label": "把…当评委", "note": "本质是让 LLM 评估另一段输出。" }, "benchmark-contamination": { "label": "会受…干扰", "note": "题目见过太多，评分就可能失真。" }, "ai-bias": { "label": "可能放大…", "note": "评判标准不稳时，偏见也会被带入。" }, "human-in-the-loop": { "label": "常配合…复核", "note": "关键场景仍需要人工兜底把关。" } } } } }, { "id": "llm-game-benchmark", "name": "LLM Game Benchmark", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "model-leaderboard" }, { "to": "agent-harness" }, { "to": "computer-use" }, { "to": "benchmark-contamination" } ], "track": "today", "i18n": { "en": { "fullName": "Large Language Model Game Benchmark", "factExplain": "A test that judges LLMs by making them play games.", "humanExplain": "Talk is cheap. Put the AI in Mario Kart and let banana peels grade it.\n\nIt checks planning and memory. It also checks if the AI can use a game screen. You often see the scores on leaderboards.", "humanExplainDisplay": "Talk is cheap.\nPut the AI in ==Mario Kart==\nand let ==banana peels== grade it.\n\nIt checks planning and memory.\nIt also checks if the AI can use a game screen.\nYou often see the scores on leaderboards.", "relationsNarrative": "Leaderboard\nGame scores often go into leaderboards to compare models.\n\nAgent harness\nMany game benchmarks use an agent harness to run tests again and again.\n\nComputer use\nGame tasks often make the model watch the screen and take actions.\n\nBenchmark contamination\nIf training data included the game tasks, the score may look too high.", "relations": { "model-leaderboard": { "label": "feeds scores to …", "note": "Game scores often get added to leaderboards." }, "agent-harness": { "label": "often runs through …", "note": "An agent harness runs many game tests the same way." }, "computer-use": { "label": "often tests … skills", "note": "Games make the model watch the screen and act." }, "benchmark-contamination": { "label": "must avoid …", "note": "Leaked game tasks can make scores look too high." } } }, "zh": { "fullName": "大语言模型游戏基准", "factExplain": "用游戏任务测试大模型能力的评测基准。", "humanExplain": "别听它嘴上挺能说，拎去打两把斗地主，记牌、配合、算后手，水平马上见光。\n\n常用来测规划、记忆和操作能力，比纯聊天题更像真干活。", "humanExplainDisplay": "别听它嘴上\n挺能说，\n拎去==打两把斗地主==，\n水平马上==见光==。\n\n常用来测规划、记忆\n和操作能力，\n比纯聊天题更像真干活。", "relationsNarrative": "Leaderboard\n游戏评测分数常被整理进排行榜比较模型强弱。\n\nAgent Harness\n很多游戏基准要靠评测框架批量运行与复现。\n\nComputer Use\n游戏任务常要求模型观察界面并执行具体操作。\n\nBenchmark Contamination\n如果训练见过题目，游戏分数就会被高估。", "relations": { "model-leaderboard": { "label": "为…提供成绩", "note": "游戏评测结果常被汇总进排行榜。" }, "agent-harness": { "label": "常由…执行", "note": "很多游戏测试要靠评测框架跑通。" }, "computer-use": { "label": "常测试…能力", "note": "游戏任务常要求观察环境并操作。" }, "benchmark-contamination": { "label": "需防…污染", "note": "题目泄露会让分数失真变好看。" } } } } }, { "id": "llm-inference-engine", "name": "Inference engine", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-05-30T03:10:23.228Z", "relations": [ { "to": "inference" }, { "to": "continuous-batching" }, { "to": "gpu" }, { "to": "vram" } ], "track": "today", "i18n": { "en": { "fullName": "LLM inference engine", "factExplain": "Software that makes a large language model generate answers efficiently.", "humanExplain": "The model is a recipe card. The inference engine is the diner cook keeping the grill busy, not staring at an empty pan.\n\nIt runs the model fast after launch. You meet it in AI APIs and local setups.", "humanExplainDisplay": "The model is a ==recipe card==.\nThe inference engine is the ==diner cook==\nkeeping the grill busy,\nnot staring at an empty pan.\n\nIt runs the model fast after launch.\nYou meet it in AI APIs\nand local setups.", "relationsNarrative": "Inference\nAn LLM inference engine organizes inference and runs it for real.\n\nContinuous batching\nContinuous batching helps the engine reduce waiting and wasted time.\n\nGPU\nThe engine schedules GPU power to make generation faster and steadier.\n\nVRAM\nVRAM limits the model size and how many requests can run at once.", "relations": { "inference": { "label": "runs … for real", "note": "It organizes inference and makes it run." }, "continuous-batching": { "label": "uses … to speed up", "note": "Continuous batching keeps requests moving with less waiting." }, "gpu": { "label": "schedules … power", "note": "The engine tries to keep the GPU busy." }, "vram": { "label": "is limited by …", "note": "VRAM size often decides the model size and traffic load." } } }, "zh": { "fullName": "大语言模型推理引擎", "factExplain": "负责高效运行大语言模型生成任务的软件系统。", "humanExplain": "推理引擎像外卖店后厨：你点一句话，它调火候、排队伍，尽快把答案炒出来。\n\n它决定模型上线后的速度、成本和稳定性，常见于聊天机器人和企业部署。", "humanExplainDisplay": "推理引擎像==外卖店后厨==：\n你点一句话，\n它调火候、排队伍，\n尽快==把答案炒出来==。\n\n它决定模型上线后的速度、\n成本和稳定性，\n常见于聊天机器人和企业部署。", "relationsNarrative": "Inference\nLLM inference engine 就是把模型推理真正组织并运行起来的软件层。\n\nContinuous batching\nContinuous batching 是推理引擎常用的提速办法，用来减少等待和空转。\n\nGPU\n推理引擎需要调度 GPU 算力，尽量让生成过程更快更稳。\n\nVRAM\nVRAM 容量会限制推理引擎能装下多大模型，以及并发开多高。", "relations": { "inference": { "label": "负责落地…", "note": "它把推理过程真正跑起来。" }, "continuous-batching": { "label": "常用…提速", "note": "连续批处理是常见优化手段。" }, "gpu": { "label": "调度…算力", "note": "推理引擎要尽量喂饱 GPU。" }, "vram": { "label": "受…容量限制", "note": "显存大小常决定能跑多大模型。" } } } } }, { "id": "llm", "name": "LLM", "aliases": [ "大模型" ], "layer": "L3", "era": "2020", "publishedAt": "2026-05-23T08:00:00Z", "relations": [ { "to": "token" }, { "to": "prompt" }, { "to": "context-window" }, { "to": "hallucination" } ], "i18n": { "en": { "fullName": "Large Language Model", "factExplain": "An AI model for understanding, writing, and rewriting language.", "humanExplain": "An LLM is the friend at lunch with a story for everything. It sounds like it read the whole internet before homeroom.\n\nYou meet it in chatbots and writing tools. It can write emails, polish text, or explain code. But it can be wrong with a straight face.", "humanExplainDisplay": "An LLM is the ==friend at lunch==\nwith a story for everything.\nIt sounds like it read the ==whole internet==\nbefore homeroom.\n\nYou meet it in chatbots and writing tools.\nIt can write emails,\npolish text,\nor explain code.\nBut it can be wrong\nwith a straight face.", "relationsNarrative": "Token\nAn LLM breaks language into Tokens, then builds its answer step by step.\n\nPrompt\nThe Prompt sets the LLM’s task and steers its answer.\n\nContext-window\nThe Context-window limits how much information the LLM can use at once.\n\nHallucination\nHallucination risk rises when the LLM lacks solid facts.", "relations": { "token": { "label": "processes text as …", "note": "An LLM breaks language into Tokens to build answers step by step." }, "prompt": { "label": "takes instructions from …", "note": "The Prompt sets the task and steers the answer." }, "context-window": { "label": "memory limit set by …", "note": "The Context-window sets how much text the LLM can use at once." }, "hallucination": { "label": "may produce …", "note": "When evidence is weak, an LLM may invent a confident answer." } } }, "zh": { "fullName": "大语言模型", "factExplain": "一种能够理解、生成和改写语言内容的 AI 模型。", "humanExplain": "大模型像家庭群里的万能亲戚，啥都能接话，还接得挺像那么回事。\n\n它能写、问答、翻译和写代码，是聊天助手、搜索、办公应用的底座。", "humanExplainDisplay": "大模型像==家庭群里的万能亲戚==，\n啥都能接话，\n还接得挺像那么回事。\n\n它能写、问答、翻译和写代码，\n是聊天助手、搜索、办公应用的底座。", "relationsNarrative": "Token\nLLM 依靠 Token 拆解语言，才能逐步生成连续回答。\n\nPrompt\nPrompt 决定 LLM 的任务边界，也影响回答的方向。\n\nContext-window\nContext-window 限制 LLM 一次能参考的信息范围。\n\nHallucination\n当 LLM 缺少可靠依据时，Hallucination 风险会明显上升。", "relations": { "token": { "label": "以…为处理单位" }, "prompt": { "label": "靠…接收指令" }, "context-window": { "label": "记忆上限由…决定" }, "hallucination": { "label": "可能产生…" } } } } }, { "id": "llmops", "name": "LLMOps", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "inference" }, { "to": "fine-tuning" }, { "to": "llm-as-a-judge" } ], "track": "today", "i18n": { "en": { "fullName": "Large Language Model Operations", "factExplain": "The set of practices for running, watching, and improving LLM systems over time.", "humanExplain": "Launching an LLM is not the finish line. LLMOps is like caring for a talking factory machine. You watch the gauges. You catch alarms fast.\n\nIt is used in real AI products at work. It keeps models live, stable, and easier to update.", "humanExplainDisplay": "Launching an LLM is not\nthe ==finish line==.\nLLMOps is like caring for\na ==talking factory machine==.\nYou watch the gauges.\nYou catch alarms fast.\n\nIt is used in real AI products at work.\nIt keeps models live, stable,\nand easier to update.", "relationsNarrative": "LLM\nLLMOps centers on getting the LLM live, watched, and updated.\n\nInference\nLLMOps works to keep inference stable in daily use.\n\nFine-tuning\nLLMOps manages new versions made by fine-tuning.\n\nLLM-as-a-judge\nLLMOps often uses LLM-as-a-judge for automatic evaluation.", "relations": { "llm": { "label": "runs ops around …", "note": "The main thing LLMOps manages is usually the LLM." }, "inference": { "label": "keeps … live", "note": "Stable inference is a big job in LLMOps." }, "fine-tuning": { "label": "manages updates from …", "note": "New fine-tuned versions also need release and monitoring." }, "llm-as-a-judge": { "label": "evaluates with …", "note": "LLMOps often uses this for automatic checks." } } }, "zh": { "fullName": "大语言模型运维", "factExplain": "管理 LLM 开发、部署、监控与迭代的一套实践。", "humanExplain": "大模型运维像管网红奶茶店：配方、排队、差评、翻车，都得有人盯着。\n\n它让模型应用可上线、可监控、可回滚，常见于客服、搜索和内部工具。", "humanExplainDisplay": "大模型运维像==管网红奶茶店==：\n配方、排队、差评、==翻车==，\n都得有人盯着。\n\n它让模型应用可上线、可监控、可回滚，\n常见于客服、搜索和内部工具。", "relationsNarrative": "LLM\nLLMOps 主要围绕 LLM 的上线、监控和迭代展开。\n\nInference\nInference 是 LLMOps 要保障稳定运行的核心环节。\n\nFine-tuning\nFine-tuning 产生的新版本，通常由 LLMOps 统一管理。\n\nLLM-as-a-judge\nLLMOps 常用 LLM-as-a-judge 做自动化评估。", "relations": { "llm": { "label": "围着…做运维", "note": "LLMOps 服务的核心对象通常就是 LLM。" }, "inference": { "label": "保障…上线", "note": "推理服务是否稳定，是 LLMOps 重点。" }, "fine-tuning": { "label": "管理…迭代", "note": "微调后的版本也要纳入发布和监控。" }, "llm-as-a-judge": { "label": "用…做评估", "note": "自动化评测常是 LLMOps 的一环。" } } } } }, { "id": "local-ai-hardware-barrier", "name": "Local HW barrier", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "local-llm" }, { "to": "gpu" }, { "to": "vram" }, { "to": "quantization" } ], "track": "today", "i18n": { "en": { "fullName": "Local AI Hardware Barrier", "factExplain": "The cost and power needed to run AI on your own device.", "humanExplain": "It is like shoving a party pizza into a toaster oven. The pizza is great, but the oven says no, so you cook slices one by one.\n\nIt decides whether a local model runs well. You meet it in home AI setups. You meet it in offline assistants. Companies meet it on private office networks.", "humanExplainDisplay": "It is like shoving a ==party pizza==\ninto a toaster oven.\nThe pizza is great,\nbut the oven says no,\nso you cook ==slices one by one==.\n\nIt decides whether a local model runs well.\nYou meet it in home AI setups.\nYou meet it in offline assistants.\nCompanies meet it on private office networks.", "relationsNarrative": "Local-LLM\nThe hardware barrier limits how many people can use Local-LLMs.\n\nGPU\nGPU power often decides whether local AI feels smooth.\n\nVRAM\nVRAM size often decides whether the model fits on your device.\n\nQuantization\nQuantization shrinks the model and lowers the hardware barrier.", "relations": { "local-llm": { "label": "limits … adoption", "note": "A local model runs only if the machine is strong enough." }, "gpu": { "label": "depends on … power", "note": "GPU power is often the first hurdle for local AI." }, "vram": { "label": "gets blocked by …", "note": "Low VRAM can stop a model before it even loads." }, "quantization": { "label": "is lowered by …", "note": "Quantization shrinks models, so weaker machines can run them." } } }, "zh": { "fullName": "本地 AI 硬件门槛", "factExplain": "指在本地运行 AI 所需硬件成本与性能门槛。", "humanExplain": "像老小区电梯搬家：你东西再多再好，门太窄、载重不够，==大件根本塞不进==，最后只能==拆箱分批搬==。\n\n它决定本地模型能否跑顺，常见于个人部署、离线助手和企业内网。", "humanExplainDisplay": "像老小区电梯搬家：\n你东西再多再好，\n门太窄、载重不够，\n==大件根本塞不进==，\n最后只能==拆箱分批搬==。\n\n它决定本地模型能否跑顺，\n常见于个人部署、\n离线助手和企业内网。", "relationsNarrative": "Local-LLM\n本地硬件门槛直接限制本地模型的可用性与普及。\n\nGPU\n显卡性能往往决定本地运行是否流畅可用。\n\nVRAM\n显存大小常决定模型能否装进本地设备。\n\nQuantization\n量化通过压缩模型，降低本地运行门槛。", "relations": { "local-llm": { "label": "限制…普及", "note": "本地模型能否跑起来，先看机器够不够。" }, "gpu": { "label": "主要看…性能", "note": "显卡算力常是本地部署的第一道坎。" }, "vram": { "label": "常被…卡住", "note": "显存不够时，模型往往连加载都难。" }, "quantization": { "label": "靠…降门槛", "note": "量化能减小体积，换来更低硬件要求。" } } } } }, { "id": "local-first-ai", "name": "Local-first AI", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "local-llm" }, { "to": "edge-ai" }, { "to": "data-privacy" }, { "to": "in-browser-ai-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Local-first AI", "factExplain": "A design that runs AI on your own device first.", "humanExplain": "Local-first AI is like making toast in your kitchen. You do not call a bakery for one slice.\n\nIt keeps small AI jobs on your phone or laptop for privacy and offline use. Big jobs can still use the cloud.", "humanExplainDisplay": "Local-first AI is like\n==making toast in your kitchen==.\nYou do not call ==a bakery==\nfor one slice.\n\nIt keeps small AI jobs on your phone or laptop\nfor privacy and offline use.\nBig jobs can still use the cloud.", "relationsNarrative": "Local-LLM\nLocal-first AI often uses a Local-LLM to run on the device.\n\nEdge AI\nEdge AI keeps the AI close to the user and the data.\n\nData-privacy\nLocal-first AI sends less data away, so privacy risk is lower.\n\nIn-browser AI\nIn-browser AI is one way to run local-first AI on a device.", "relations": { "local-llm": { "label": "runs with …", "note": "A Local-LLM is a common way to keep AI on the device." }, "edge-ai": { "label": "stays near devices with …", "note": "Edge devices bring AI closer to the user." }, "data-privacy": { "label": "sends less data out for …", "note": "Less data leaves the device, so privacy risk is lower." }, "in-browser-ai-ai": { "label": "works through …", "note": "A browser can keep AI work on the same machine." } } }, "zh": { "fullName": "本地优先 AI", "factExplain": "优先在本地设备运行 AI 的设计方式。", "humanExplain": "本地优先 AI 像小区煎饼摊：大清早能热乎现做，就别为加个蛋跑中央厨房。\n\n适合隐私、离线场景；重活再交给云端。", "humanExplainDisplay": "本地优先 AI 像\n小区煎饼摊：\n大清早能热乎现做，\n别为加个蛋==跑中央厨房==。\n\n适合隐私、离线场景；\n重活再交给云端。", "relationsNarrative": "Local-LLM\n本地语言模型是它最常见的实现方式。\n\nEdge AI\n边缘设备让它离用户和数据更近。\n\nData-privacy\n它让数据少出门，降低外传风险。\n\nIn-browser AI\n浏览器内运行，是本地优先的一种入口。", "relations": { "local-llm": { "label": "用…本地运行", "note": "小模型本地跑，是常见落地形态。" }, "edge-ai": { "label": "依托…靠近设备", "note": "边缘设备让 AI 离用户更近。" }, "data-privacy": { "label": "减少…外传", "note": "数据少出门，隐私风险更低。" }, "in-browser-ai-ai": { "label": "通过…落地", "note": "浏览器可把能力留在本机。" } } } } }, { "id": "local-llm", "name": "Local-LLM", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-05-23T10:45:00Z", "relations": [ { "to": "open-source-model" }, { "to": "llm" }, { "to": "gpu" }, { "to": "data-privacy" } ], "i18n": { "en": { "fullName": "Local Large Language Model", "factExplain": "A large language model running on your own device or private server.", "humanExplain": "A Local LLM is AI living in your spare room. It keeps secrets well, but it may hog the power bill.\n\nIt helps with offline work and private files. It can answer fast, but hardware sets the limit.", "humanExplainDisplay": "A Local LLM is AI\nliving in your ==spare room==.\nIt keeps ==secrets== well,\nbut it may hog the power bill.\n\nIt helps with offline work\nand private files.\nIt can answer fast,\nbut hardware sets the limit.", "relationsNarrative": "Open-source-model\nOpen-source models give Local LLMs something to download and run.\n\nLLM\nA Local LLM is an LLM running on a device or private server.\n\nGPU\nThe GPU affects how large and fast a Local LLM can be.\n\nData-privacy\nData privacy is a key reason people choose a Local LLM.", "relations": { "open-source-model": { "label": "often uses …", "note": "Open-source models give Local LLMs a base to download and run." }, "llm": { "label": "runs … locally", "note": "A Local LLM is an LLM deployed on a device or private server." }, "gpu": { "label": "may need …", "note": "The GPU affects model size and speed." }, "data-privacy": { "label": "supports … needs", "note": "Data privacy is a key reason to choose a Local LLM." } } }, "zh": { "fullName": "本地大模型", "factExplain": "在本地设备或私有环境中运行的大语言模型。", "humanExplain": "它像把 AI 师傅请到家里干活，不用排云端长队，悄悄话也少外传。\n\n它适合隐私文档、离线办公和开发调试，但很吃电脑配置。", "humanExplainDisplay": "它像==把 AI 师傅请到家里干活==，\n不用==排云端长队==，\n悄悄话也少外传。\n\n它适合隐私文档、\n离线办公和开发调试，\n但很吃电脑配置。", "relationsNarrative": "Open-source-model\nOpen-source-model 为 Local-LLM 提供可下载和部署的基础。\n\nLLM\nLocal-LLM 是将 LLM 部署在本地或私有环境中。\n\nGPU\nGPU 决定 Local-LLM 可运行的模型规模和速度。\n\nData-privacy\nData-privacy 是选择 Local-LLM 的重要原因之一。", "relations": { "open-source-model": { "label": "依赖…" }, "gpu": { "label": "可能需要…" }, "data-privacy": { "label": "服务于…场景" } } } } }, { "id": "logic-programming", "name": "Logic", "layer": "L1", "era": "1972", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "ontology" }, { "to": "description-logic" }, { "to": "constraint-satisfaction-problem" } ], "track": "history", "i18n": { "en": { "fullName": "Logic Programming", "factExplain": "A way to solve problems using facts, rules, and reasoning.", "humanExplain": "Logic programming is like writing rules for a board game. The computer becomes the referee with a tiny whistle.\n\nYou write facts and if-then rules. It checks cases and works out the answer.", "humanExplainDisplay": "Logic programming is like writing ==rules for a board game==.\nThe computer becomes the ==referee==\nwith a tiny whistle.\n\nYou write facts and if-then rules.\nIt checks cases\nand works out the answer.", "relationsNarrative": "KR\nLogic programming writes knowledge as facts and rules a machine can reason with.\n\nOntology\nOntology sets the ideas and links first. Logic programming then reasons with them.\n\nDL\nLogic programming and DL both use symbols and strict rules for reasoning.\n\nCSP\nMany CSPs can be written as rules. Logic programming can then find an answer.", "relations": { "knowledge-representation": { "label": "implements …", "note": "It writes knowledge as facts and rules." }, "ontology": { "label": "models with …", "note": "Ontology sets the ideas and links first." }, "description-logic": { "label": "shares roots with …", "note": "Both use symbols and strict rules for reasoning." }, "constraint-satisfaction-problem": { "label": "can solve …", "note": "Write the limits clearly, then let it find a fit." } } }, "zh": { "fullName": "Logic Programming（逻辑编程）", "factExplain": "用事实和规则描述问题并靠推理求解的范式。", "humanExplain": "别一行行教它，像给桌游写规则书：谁能走、抽到什么牌怎么应对、碰上冲突算谁赢，它当裁判照着条文自己判。\n\n适合把条件逻辑写成程序，常用于规则判断和自动推理。", "humanExplainDisplay": "别一行行教它，\n像给桌游写==规则书==：\n谁能走、抽到什么牌怎么应对、\n碰上冲突算谁赢，\n它当==裁判==照着条文自己判。\n\n适合把条件逻辑\n写成程序，\n常用于规则判断\n和自动推理。", "relationsNarrative": "Knowledge Representation\n逻辑编程常被用来把知识写成可推理的规则与事实。\n\nOntology\n本体先定义概念和关系，逻辑编程再据此做推理。\n\nDescription Logic\n两者都属于符号主义路线，强调形式化与可推理。\n\nConstraint-satisfaction-problem\n很多约束问题可改写成规则，再由它自动求解。", "relations": { "knowledge-representation": { "label": "实现…的一种方式", "note": "常用规则和事实表达知识。" }, "ontology": { "label": "常配合…建模", "note": "本体先定概念边界与关系。" }, "description-logic": { "label": "与…同属符号派", "note": "都重视形式规则与可推理性。" }, "constraint-satisfaction-problem": { "label": "可用于求解…", "note": "把约束写清后再自动找解。" } } } } }, { "id": "logic-theorist", "name": "Logic Theorist", "layer": "L4", "era": "1956", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "automated-theorem-proving" }, { "to": "symbolic-ai" }, { "to": "dartmouth-workshop" }, { "to": "general-problem-solver" } ], "track": "history", "i18n": { "en": { "fullName": "Logic Theorist", "factExplain": "A 1956 AI program for proving math theorems automatically.", "humanExplain": "Logic Theorist was the kid in math class with the answer already raised. The chalk squeaked once, and it was halfway through the proof.\n\nIt searched for proof steps by itself. People used it to prove math theorems.", "humanExplainDisplay": "Logic Theorist was the ==kid in math class==\nwith the answer already raised.\nThe chalk squeaked once,\nand it was ==halfway through the proof==.\n\nIt searched for proof steps by itself.\nPeople used it to prove math theorems.", "relationsNarrative": "ATP\nLogic Theorist was an early model for proving math theorems by machine.\n\nSymbolic AI\nLogic Theorist used symbols, rules, and search to show early Symbolic AI.\n\nDartmouth Workshop\nLogic Theorist showed the promise of machine reasoning at the Dartmouth Workshop.\n\nGPS\nGPS kept its idea of breaking problems into searchable steps.", "relations": { "automated-theorem-proving": { "label": "helped start …", "note": "It was an early model for proving theorems by machine." }, "symbolic-ai": { "label": "showed …", "note": "It used symbols and rules to imitate reasoning." }, "dartmouth-workshop": { "label": "made waves at …", "note": "It became famous around the Dartmouth Workshop." }, "general-problem-solver": { "label": "inspired …", "note": "GPS used a similar search idea." } } }, "zh": { "fullName": "逻辑理论家", "factExplain": "1956 年用于自动证明数学定理的早期 AI 程序。", "humanExplain": "Logic Theorist 是数学课抢答王：题刚上黑板，它先把证明写完。\n\n它能自动搜索证明步骤，用在数学定理证明场景。", "humanExplainDisplay": "Logic Theorist 是\n数学课==抢答王==：\n题刚上黑板，\n它先把证明写完。\n\n它能自动搜索证明步骤，\n用在数学定理证明场景。", "relationsNarrative": "Automated Theorem Proving\n它是自动证明数学定理的早期样板。\n\nSymbolic AI\n它用符号、规则和搜索展示早期符号派思路。\n\nDartmouth Workshop\n它在达特茅斯会议上展示了机器推理潜力。\n\nGeneral Problem Solver\nGPS 延续了它把问题拆开搜索的思路。", "relations": { "automated-theorem-proving": { "label": "开创…", "note": "它是自动定理证明的早期样板。" }, "symbolic-ai": { "label": "代表…", "note": "它用符号和规则模拟推理。" }, "dartmouth-workshop": { "label": "亮相于…", "note": "它在达特茅斯会议前后成名。" }, "general-problem-solver": { "label": "启发…", "note": "GPS 延续了类似的搜索思路。" } } } } }, { "id": "logistic-regression", "name": "Logit", "layer": "L2", "era": "1958", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "supervised-learning" }, { "to": "cross-entropy-loss" }, { "to": "gradient-descent" } ], "track": "history", "i18n": { "en": { "fullName": "Logistic Regression", "factExplain": "A linear supervised model that predicts the probability of one of two classes.", "humanExplain": "Logit is a nightclub bouncer with a tiny calculator. It adds your score, then says “in” or “not tonight.”\n\nYou meet it in risk checks, health screens, and click guesses. It works best when the yes/no line is clear.", "humanExplainDisplay": "Logit is a ==nightclub bouncer==\nwith a tiny calculator.\nIt adds your score,\nthen says ==“in” or “not tonight.”==\n\nYou meet it in risk checks,\nhealth screens,\nand click guesses.\nIt works best when\nthe yes/no line is clear.", "relationsNarrative": "Classification\nLogit is one of the classic methods for two-class decisions.\n\nSupervised Learning\nLogit uses labeled examples to learn a dividing line.\n\nCross-Entropy Loss\nLogit often uses Cross-Entropy Loss to measure prediction errors.\n\nGradient Descent\nLogit's weights can be learned with Gradient Descent.", "relations": { "classification": { "label": "is a … method", "note": "It is one of the classic two-class models." }, "supervised-learning": { "label": "learns through …", "note": "It learns the dividing line from labeled examples." }, "cross-entropy-loss": { "label": "often trains with …", "note": "Cross-Entropy Loss measures how wrong its guesses are." }, "gradient-descent": { "label": "can be solved by …", "note": "Gradient Descent updates its weights during training." } } }, "zh": { "fullName": "逻辑回归", "factExplain": "一种用概率做二分类的线性监督学习方法。", "humanExplain": "它像小区门禁看访客码：先按几项信息算分，过线放行，不过线就拦在门外。\n\n常用于风控、医疗初筛和点击预测，适合做边界清楚的二分类判断。", "humanExplainDisplay": "它像小区门禁看访客码：\n先按几项信息==算分==，\n过线放行，\n不过线就==拦在门外==。\n\n常用于风控、医疗初筛\n和点击预测，\n适合做边界清楚的二分类判断。", "relationsNarrative": "Classification\n它是最经典的二分类方法之一。\n\nSupervised Learning\n它依赖带标签样本学习分类边界。\n\nCross-Entropy Loss\n它常用交叉熵来衡量预测误差。\n\nGradient Descent\n它的参数通常可用梯度下降求解。", "relations": { "classification": { "label": "属于…方法", "note": "它是最经典的二分类模型之一。" }, "supervised-learning": { "label": "依赖…训练", "note": "它靠带标签数据学出分界线。" }, "cross-entropy-loss": { "label": "常配合…优化", "note": "训练时常用它衡量预测误差。" }, "gradient-descent": { "label": "可用…求解", "note": "参数通常靠梯度更新来学习。" } } } } }, { "id": "long-horizon-task", "name": "Long-horizon", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "agent-memory" }, { "to": "computer-use" }, { "to": "reasoning-model" } ], "track": "today", "i18n": { "en": { "fullName": "Long-horizon task", "factExplain": "A task that needs many planned steps, steady work, and fixes along the way.", "humanExplain": "It is like running the school play. The cast list is only the start. Then the dragon costume vanishes.\n\nYou see it in coding. You also see it in workflows and software work. When the steps get long, AI can skip parts or wander off.", "humanExplainDisplay": "It is like running the ==school play==.\nThe cast list is only the start.\nThen the ==dragon costume vanishes==.\n\nYou see it in coding.\nYou also see it in workflows\nand software work.\nWhen the steps get long,\nAI can skip parts or wander off.", "relationsNarrative": "Agent\nMany Agent systems are built to finish this kind of multi-step task.\n\nMemory\nLong-horizon tasks need Memory, or the AI loses the thread.\n\nComputer use\nReal software and websites often turn work into a long-horizon task.\n\nReasoning-model\nA Reasoning-model can help split the steps, but long paths can still drift.", "relations": { "agent": { "label": "often a goal for …", "note": "Many Agent systems are built to finish these multi-step tasks." }, "agent-memory": { "label": "leans on …", "note": "Long tasks need Memory so the AI does not lose the thread." }, "computer-use": { "label": "often lands through …", "note": "Real apps and screens can turn work into a long task." }, "reasoning-model": { "label": "tests … planning", "note": "Long tasks need plans, plus fixes when the path changes." } } }, "zh": { "fullName": "长程任务", "factExplain": "需要多步规划、长期执行与持续纠错的任务。", "humanExplain": "这活儿像带班排毕业晚会，节目单定了才开头，后面还得一轮轮催人、补位、救场，不能演到压轴才发现串词拿错。\n\n常见于写代码、跑流程、操作软件；步骤一长，就容易漏环节和跑偏。", "humanExplainDisplay": "这活儿像带班排\n==毕业晚会==，\n后面还得一轮轮催人、\n==补位救场==。\n\n常见于写代码、跑流程，\n操作软件；\n步骤一长，\n就容易漏环节和跑偏。", "relationsNarrative": "Agent\n很多代理系统的核心目标，就是完成这类多步任务。\n\nMemory\n长程任务很吃记忆，否则做着做着就断片。\n\nComputer use\n一旦要操作真实软件和网页，常会变成长程任务。\n\nReasoning-model\n推理模型能帮它拆步骤，但链路长了仍会跑偏。", "relations": { "agent": { "label": "常作为…目标", "note": "很多代理系统就是为这类任务设计。" }, "agent-memory": { "label": "依赖…续航", "note": "步骤一长，记忆能力就变关键。" }, "computer-use": { "label": "常借…落地", "note": "要碰真实界面时更像长程任务。" }, "reasoning-model": { "label": "考验…规划", "note": "任务越长，越需要中途修正路线。" } } } } }, { "id": "lora", "name": "LoRA", "layer": "L2", "era": "2021", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "fine-tuning" }, { "to": "parameter" }, { "to": "quantization" }, { "to": "local-llm" } ], "track": "history", "i18n": { "en": { "fullName": "Low-Rank Adaptation", "factExplain": "A fine-tuning method using only a small set of new parameters.", "humanExplain": "LoRA is like a clip-on cup holder for a bike. No bike surgery. Your latte gets a safer ride.\n\nIt lets teams teach a model new work tricks with less memory. You meet it in custom company models and local AI setups.", "humanExplainDisplay": "LoRA is like a ==clip-on cup holder== for a bike.\nNo ==bike surgery==.\nYour latte gets a safer ride.\n\nIt lets teams teach a model new work tricks\nwith less memory.\nYou meet it in custom company models\nand local AI setups.", "relationsNarrative": "Fine-tuning\nLoRA is a light Fine-tuning method with few trained parameters.\n\nParameter\nLoRA freezes most Parameters and trains small new ones.\n\nQuantization\nLoRA often pairs with Quantization to cut memory and cost.\n\nLocal-LLM\nLoRA makes Local-LLM changes more realistic on small machines.", "relations": { "fine-tuning": { "label": "is common in …", "note": "LoRA is a cheaper fine-tuning path." }, "parameter": { "label": "changes few …", "note": "It freezes most old parameters and trains small new ones." }, "quantization": { "label": "pairs with …", "note": "Together they cut memory needs and setup cost." }, "local-llm": { "label": "lowers the bar for …", "note": "It makes model tweaks easier on personal devices." } } }, "zh": { "fullName": "低秩适配", "factExplain": "一种只训练少量新增参数的微调方法。", "humanExplain": "LoRA 像给手机换个轻薄壳加插件：主板不用大修，花小钱小内存，也能立刻多出新本事。\n\n常用于低成本微调，适合行业定制和本地部署。", "humanExplainDisplay": "LoRA 像给手机换个==轻薄壳加插件==：\n主板不用大修，\n花小钱小内存，\n也能立刻多出==新本事==。\n\n常用于低成本微调，\n适合行业定制和本地部署。", "relationsNarrative": "Fine-tuning\nLoRA 是微调里常见的轻量方案，主打少改参数。\n\nParameter\n它通常冻结原参数，只训练额外加入的小部分参数。\n\nQuantization\nLoRA 常和量化一起用，进一步压低显存和成本。\n\nLocal-LLM\n它让本地模型定制更现实，不必动辄全量重训。", "relations": { "fine-tuning": { "label": "属于…常用做法", "note": "它是更省资源的微调路线。" }, "parameter": { "label": "少改…来适配", "note": "核心是冻结大部分原参数。" }, "quantization": { "label": "可和…搭配用", "note": "两者常一起压低部署门槛。" }, "local-llm": { "label": "帮…降门槛", "note": "让个人设备也更容易改模型。" } } } } }, { "id": "lstm", "name": "LSTM", "layer": "L3", "era": "1997", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "seq2seq" }, { "to": "transformer" }, { "to": "speech-to-text" } ], "track": "history", "i18n": { "en": { "fullName": "Long Short-Term Memory", "factExplain": "A recurrent neural network for step-by-step information, like speech or text.", "humanExplain": "LSTM is like a friend watching a mystery show with you. It ignores the chip ads, but remembers the muddy shoe for episode ten.\n\nIt was used for speech and time-series data. In early text AI, it helped early clues matter later.", "humanExplainDisplay": "LSTM is like a friend\nwatching a ==mystery show== with you.\nIt ignores the chip ads,\nbut remembers the ==muddy shoe==\nfor episode ten.\n\nIt was used for speech\nand time-series data.\nIn early text AI,\nit helped early clues matter later.", "relationsNarrative": "Neural-network\nLSTM is a classic Neural-network member for ordered data.\n\nSeq2Seq\nEarly Seq2Seq often used LSTM to encode and decode sequences.\n\nTransformer\nTransformer later replaced LSTM in many text tasks.\n\nSTT\nLSTM was used for years in STT and other time-based tasks.", "relations": { "neural-network": { "label": "is a branch of …", "note": "It is a classic neural network for ordered data." }, "seq2seq": { "label": "was often used by …", "note": "Early Seq2Seq used LSTMs to read and write sequences." }, "transformer": { "label": "was later replaced by …", "note": "Many language tasks later moved from LSTM to Transformer." }, "speech-to-text": { "label": "once powered …", "note": "It was long used for speech recognition and other time-based tasks." } } }, "zh": { "fullName": "长短期记忆网络", "factExplain": "一种擅长处理序列信息的循环神经网络。", "humanExplain": "LSTM 像老中医问诊：无关症状先略过，关键脉象记心里，看到后面几句，前面的线索还接得上。\n\n它常用于语音、时间序列和早期文本任务，处理前后信息会互相影响的场景。", "humanExplainDisplay": "LSTM 像老中医问诊：\n无关症状先略过，\n关键脉象==记心里==，\n看到后面几句，\n前面的线索还==接得上==。\n\n它常用于语音、\n时间序列和早期文本任务，\n处理前后信息会互相影响的场景。", "relationsNarrative": "Neural-network\n它是神经网络家族里处理序列的经典成员。\n\nSeq2Seq\n早期 Seq2Seq 常用它来编码和生成序列。\n\nTransformer\n在很多文本任务上，它后来被 Transformer 逐步替代。\n\nSpeech-to-text\n它曾长期用于语音识别这类时序建模任务。", "relations": { "neural-network": { "label": "属于…分支", "note": "它是神经网络里的经典序列模型。" }, "seq2seq": { "label": "常被…采用", "note": "早期 Seq2Seq 常用它做编码解码。" }, "transformer": { "label": "被…逐步替代", "note": "很多语言任务后来转向 Transformer。" }, "speech-to-text": { "label": "曾支撑…", "note": "它长期用于语音识别等时序任务。" } } } } }, { "id": "maas-model-as-a-service", "name": "MaaS", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "api" }, { "to": "llm" }, { "to": "on-premise-ai" }, { "to": "llmops" } ], "track": "today", "i18n": { "en": { "fullName": "Model as a Service", "factExplain": "A way to use AI models through cloud APIs as a service.", "humanExplain": "MaaS is a food delivery app for AI models. No oven, no chef hat, just tap and use.\n\nIt helps apps add AI fast. You skip setup and upkeep, but rely more on the cloud provider.", "humanExplainDisplay": "MaaS is a ==food delivery app== for AI models.\nNo oven, no chef hat,\njust ==tap and use==.\n\nIt helps apps add AI fast.\nYou skip setup and upkeep,\nbut rely more on the cloud provider.", "relationsNarrative": "API\nMaaS usually offers models through an API, so apps can plug in fast.\n\nLLM\nMany MaaS products are LLM services you can call when needed.\n\nOn-premise AI\nMaaS is the cloud path, while On-premise AI is the local path.\n\nLLMOps\nLLMOps helps teams monitor and manage many model services.", "relations": { "api": { "label": "usually arrives through …", "note": "A cloud API is the usual front door for MaaS." }, "llm": { "label": "wraps … as a service", "note": "Many cloud providers sell LLM access as a ready service." }, "on-premise-ai": { "label": "contrasts with …", "note": "MaaS runs in the cloud, while On-premise AI runs locally." }, "llmops": { "label": "is managed by …", "note": "LLMOps helps track model calls, health, and cost." } } }, "zh": { "fullName": "模型即服务", "factExplain": "把模型通过云端接口按服务方式提供。", "humanExplain": "做 AI 不必自己搭灶台了，MaaS 更像楼下便利店：缺什么模型，随手拿了就能结账开用。\n\n适合快速接入 AI 功能，省部署运维，但更依赖云厂商。", "humanExplainDisplay": "做 AI 不必自己搭灶台了，\nMaaS 更像==楼下便利店==：\n缺什么模型，\n随手拿了就能==结账开用==。\n\n适合快速接入 AI 功能，\n省部署运维，但更依赖云厂商。", "relationsNarrative": "API\n它通常通过 API 对外提供，方便应用直接接入。\n\nLLM\n很多 MaaS 的核心商品，就是可调用的大模型服务。\n\nOn-premise AI\n它和本地部署相对：一个省事，一个更可控。\n\nLLMOps\n团队接入多个模型服务后，常用它做监控与管理。", "relations": { "api": { "label": "通常经…交付", "note": "它最常见的入口就是云接口。" }, "llm": { "label": "把…封装成服务", "note": "很多云厂商把大模型直接对外提供。" }, "on-premise-ai": { "label": "与…相对", "note": "一个走云上托管，一个走本地自建。" }, "llmops": { "label": "被…管理", "note": "多模型调用、监控和成本常归它管。" } } } } }, { "id": "machine-translation", "name": "MT", "layer": "L4", "era": "1954", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "neural-machine-translation" }, { "to": "seq2seq" }, { "to": "transformer" }, { "to": "multilingual-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Machine Translation", "factExplain": "Software that automatically turns text from one language into another.", "humanExplain": "Machine translation is like a pocket phrasebook with batteries. It can save you at a taco truck, then face-plant on a joke.\n\nIt turns text into another language. You meet it on websites, documents, and subtitles, but people should check it.", "humanExplainDisplay": "Machine translation is like\n==a pocket phrasebook with batteries==.\nIt can save you at a taco truck,\nthen ==face-plant on a joke==.\n\nIt turns text into another language.\nYou meet it on websites, documents, and subtitles,\nbut people should check it.", "relationsNarrative": "NMT\nNMT used neural networks to greatly improve machine translation quality.\n\nSeq2Seq\nSeq2Seq was a core design for early neural translation.\n\nTransformer\nTransformers made long-sentence translation steadier and faster.\n\nMultilingual AI\nMultilingual AI often uses translation as a basic skill.", "relations": { "neural-machine-translation": { "label": "grew into …", "note": "Neural networks made machine translation smoother and more natural." }, "seq2seq": { "label": "often used …", "note": "Seq2Seq reads the source sentence, then writes the target sentence." }, "transformer": { "label": "improves with …", "note": "Transformers became the main engine for modern translation." }, "multilingual-ai": { "label": "supports …", "note": "Translation is a basic skill for multilingual AI." } } }, "zh": { "fullName": "机器翻译", "factExplain": "把一种语言文本自动转换成另一种语言。", "humanExplain": "机器翻译像随身带个塑料外教：菜单能救命，碰上梗就开始装傻。\n\n用于网页、文档和字幕，帮跨语言沟通但需校对。", "humanExplainDisplay": "机器翻译像随身带个\n==塑料外教==：\n菜单能救命，\n碰上梗就开始装傻。\n\n用于网页、文档和字幕，\n帮跨语言沟通，\n但需校对。", "relationsNarrative": "NMT\nNMT 用神经网络大幅提升机器翻译质量。\n\nSeq2Seq\nSeq2Seq 是早期神经翻译的核心框架。\n\nTransformer\nTransformer 让长句翻译更稳、更并行。\n\nMultilingual AI\n多语言 AI 常把翻译作为基础能力。", "relations": { "neural-machine-translation": { "label": "发展为…", "note": "神经网络让翻译更流畅自然。" }, "seq2seq": { "label": "常用…建模", "note": "它把源句编码后生成目标句。" }, "transformer": { "label": "借助…提升质量", "note": "注意力架构成了现代翻译主力。" }, "multilingual-ai": { "label": "支撑…能力", "note": "跨语言能力离不开翻译任务。" } } } } }, { "id": "markov-chain-monte-carlo", "name": "MCMC", "layer": "L2", "era": "1953", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "variational-inference" }, { "to": "hidden-markov-model" }, { "to": "monte-carlo-tree-search" } ], "track": "history", "i18n": { "en": { "fullName": "Markov Chain Monte Carlo", "factExplain": "A sampling method for estimating hard probability patterns with many random steps.", "humanExplain": "MCMC is like judging a giant school bake sale without seeing the map. You wander table to table, and your crumbs slowly reveal the crowd favorites.\n\nYou meet it in Bayesian inference. It keeps sampling until a messy probability shape becomes clearer.", "humanExplainDisplay": "MCMC is like judging a ==giant school bake sale==\nwithout seeing the map.\nYou wander table to table,\nand your ==crumbs== slowly reveal the crowd favorites.\n\nYou meet it in Bayesian inference.\nIt keeps sampling until a messy probability shape\nbecomes clearer.", "relationsNarrative": "VI\nMCMC is often compared with VI: one samples, and one uses an approximation.\n\nHMM\nMCMC can sample and estimate hidden states in an HMM.\n\nMCTS\nMCMC and MCTS both use random sampling, but MCMC maps distributions and MCTS guides choices.", "relations": { "variational-inference": { "label": "often compared with …", "note": "Both try to approximate hard probability distributions." }, "hidden-markov-model": { "label": "can estimate …", "note": "MCMC can sample hidden states in an HMM." }, "monte-carlo-tree-search": { "label": "shares sampling with …", "note": "Both use random samples to explore possibilities." } } }, "zh": { "fullName": "马尔可夫链蒙特卡罗", "factExplain": "通过构造随机采样过程来逼近复杂概率分布的方法。", "humanExplain": "像租房不看全城房源，你这小区转一圈、那条街挪一步，跑久了也能摸清哪片最适合住。\n\n它用于贝叶斯推断等场景，靠反复采样逼近复杂分布。", "humanExplainDisplay": "像租房不看全城房源，\n你这小区转一圈、\n那条街==挪一步==，\n跑久了也能摸清\n哪片最==适合住==。\n\n它用于贝叶斯推断\n等场景，\n靠反复采样逼近复杂分布。", "relationsNarrative": "Variational Inference\n它常和变分推断对比：一个靠采样，一个靠近似。\n\nHidden Markov Model\n它可用于这类含隐变量模型的后验采样与估计。\n\nMonte Carlo Tree Search\n两者都借助随机采样，但一个重分布，一个重决策。", "relations": { "variational-inference": { "label": "常与…对比", "note": "两者都在逼近难算的概率分布。" }, "hidden-markov-model": { "label": "可用于估计…", "note": "可为含隐状态的概率模型做采样推断。" }, "monte-carlo-tree-search": { "label": "共享采样思想", "note": "两者都借助随机采样探索可能性。" } } } } }, { "id": "markov-decision-process", "name": "MDP", "layer": "L2", "era": "1957", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "bellman-equation" }, { "to": "policy-gradient" }, { "to": "q-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Markov Decision Process", "factExplain": "A way to model choices, results, and rewards over time.", "humanExplain": "MDP is like picking moves in a video game level. Grab the shiny coin now, or dodge the lava so future-you survives.\n\nIt tracks each situation and each move. It is the usual setup for RL and robot control.", "humanExplainDisplay": "MDP is like picking moves\nin a ==video game level==.\nGrab the shiny coin now,\nor dodge the lava\nso ==future-you survives==.\n\nIt tracks each situation and each move.\nIt is the usual setup for RL and robot control.", "relationsNarrative": "RL\nRL often turns the environment and goal into an MDP.\n\nBellman Equation\nBellman Eq describes how value moves through an MDP.\n\nPolicy Gradient\nPolicy Gradient optimizes a policy inside an MDP.\n\nQ-Learning\nQ-Learning uses an MDP, then learns the value of each action.", "relations": { "reinforcement-learning": { "label": "sets up …", "note": "RL often writes the task as an MDP first." }, "bellman-equation": { "label": "supports …", "note": "Bellman Eq describes value from one step to the next." }, "policy-gradient": { "label": "defines the world for …", "note": "Policy Gradient usually improves a policy inside an MDP." }, "q-learning": { "label": "frames the problem for …", "note": "Q-Learning learns action values inside an MDP." } } }, "zh": { "fullName": "Markov Decision Process（马尔可夫决策过程）", "factExplain": "描述在状态中行动并获得回报的决策框架。", "humanExplain": "像下棋走子：你这一步不只图眼前吃子，还得算清会把后面整盘带到哪条路上。\n\n常用来描述强化学习任务，适合机器人控制等连续决策。", "humanExplainDisplay": "像下棋走子：\n你这一步不只图==眼前吃子==，\n还得算清会把后面整盘\n带到哪条==路上==。\n\n常用来描述强化学习任务，\n适合机器人控制等连续决策。", "relationsNarrative": "Reinforcement-learning\n强化学习通常先把环境和目标表述成这个框架。\n\nBellman-equation\n贝尔曼方程用来描述它里面的价值递推关系。\n\nPolicy-gradient\n策略梯度方法通常在这个框架下优化策略。\n\nQ-Learning\nQ 学习把决策问题写成它后再学习动作价值。", "relations": { "reinforcement-learning": { "label": "作为…基础框架", "note": "强化学习常把任务先写成它。" }, "bellman-equation": { "label": "用…求解价值", "note": "价值递推关系建立在它之上。" }, "policy-gradient": { "label": "为…定义环境", "note": "策略优化通常默认任务满足它。" }, "q-learning": { "label": "被…当作问题形式", "note": "Q 学习就是在这种框架里找最优策略。" } } } } }, { "id": "mask-r-cnn", "name": "Mask R-CNN", "layer": "L3", "era": "2017", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "faster-r-cnn" }, { "to": "object-detection" }, { "to": "image-segmentation" }, { "to": "cnn" } ], "track": "history", "i18n": { "en": { "fullName": "Mask Region-based Convolutional Neural Network", "factExplain": "A vision model for finding objects and outlining each one pixel by pixel.", "humanExplain": "Mask R-CNN is like magic scissors for a photo. It does not just box the dog. It cuts around the ears and tail.\n\nYou meet it in photo cutouts and self-driving cars. Doctors use it in scans too. It can find shapes down to pixels.", "humanExplainDisplay": "Mask R-CNN is like ==magic scissors== for a photo.\nIt does not just box the dog.\nIt cuts around the ==ears and tail==.\n\nYou meet it in photo cutouts\nand self-driving cars.\nDoctors use it in scans too.\nIt can find shapes down to pixels.", "relationsNarrative": "Faster R-CNN\nMask R-CNN adds mask prediction on top of Faster R-CNN boxes.\n\nObject Detection\nObject Detection finds object positions. Mask R-CNN also draws the outlines.\n\nImage Segmentation\nMask R-CNN is a classic way to split each object into pixels.\n\nCNN\nA CNN gives Mask R-CNN the image features it uses.", "relations": { "faster-r-cnn": { "label": "adds masks to …", "note": "It keeps Faster R-CNN boxes, then predicts each object outline." }, "object-detection": { "label": "extends …", "note": "Object Detection gives boxes. Mask R-CNN also gives outlines." }, "image-segmentation": { "label": "does …", "note": "It marks the pixels for each object." }, "cnn": { "label": "uses … for features", "note": "A CNN extracts the image features it needs." } } }, "zh": { "fullName": "掩码区域卷积神经网络", "factExplain": "一种同时做目标检测和实例分割的视觉模型。", "humanExplain": "Mask R-CNN 像裁缝量体：不只报你站哪儿，还沿肩腰腿把轮廓描到针脚。\n\n用于抠图、驾驶和医学影像，定位细到像素。", "humanExplainDisplay": "Mask R-CNN 像裁缝量体：\n不只报你==站哪儿==，\n还沿肩腰腿，\n把轮廓==描到针脚==。\n\n用于抠图、驾驶和医学影像，\n定位，\n细到像素。", "relationsNarrative": "Faster R-CNN\n它在 Faster R-CNN 检测框基础上增加掩码预测。\n\nObject Detection\n目标检测负责找出物体位置，它进一步描出轮廓。\n\nImage Segmentation\n它是实例分割的经典方法，能区分每个物体。\n\nCNN\nCNN 为它提供图像特征，是早期实现的底座。", "relations": { "faster-r-cnn": { "label": "在…上加掩码", "note": "它继承检测框，再预测物体轮廓。" }, "object-detection": { "label": "扩展…", "note": "检测只给框，它还给轮廓。" }, "image-segmentation": { "label": "实现…", "note": "它把目标分到像素级区域。" }, "cnn": { "label": "依赖…提特征", "note": "卷积网络负责提取图像特征。" } } } } }, { "id": "masked-language-modeling", "name": "MLM", "layer": "L2", "era": "2018", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "bert" }, { "to": "pretraining" }, { "to": "self-supervised-learning" }, { "to": "language-modeling" } ], "track": "history", "i18n": { "en": { "fullName": "Masked Language Modeling", "factExplain": "A training method for learning language by guessing hidden words.", "humanExplain": "MLM is like a worksheet with sticky notes over some words. The AI reads around each blank and fills in the missing word.\n\nIt is common in pretraining. It helps models get good at understanding text.", "humanExplainDisplay": "MLM is like a worksheet\nwith ==sticky notes== over some words.\nThe AI reads around each blank\nand ==fills in the missing word==.\n\nIt is common in pretraining.\nIt helps models get good at understanding text.", "relationsNarrative": "BERT\nBERT became known for using MLM to understand context.\n\nPretraining\nMLM is often used in pretraining with huge piles of text.\n\nSSL\nMLM makes questions from the text itself, so it is SSL.\n\nLM\nMLM is a type of LM, but it does not only guess the next word.", "relations": { "bert": { "label": "trains …", "note": "BERT used MLM to learn context." }, "pretraining": { "label": "is used in …", "note": "MLM is often used during pretraining." }, "self-supervised-learning": { "label": "is a kind of …", "note": "MLM makes its own quiz from text, so it needs no human labels." }, "language-modeling": { "label": "is a form of …", "note": "MLM is one way to train a language model." } } }, "zh": { "fullName": "Masked Language Modeling｜掩码语言模型", "factExplain": "一种通过预测被遮住词语来学习语言表示的训练方法。", "humanExplain": "最经典的玩法，就是把句子里几块瓷砖先抠走，让模型照着前后文把坑一点点补平。\n\n常用于语言模型预训练，尤其适合理解类任务。", "humanExplainDisplay": "最经典的玩法，\n就是把句子里几块瓷砖\n先==抠走==，\n让模型照着前后文\n把坑==一点点补平==。\n\n常用于语言模型预训练，\n尤其适合理解类任务。", "relationsNarrative": "Bert\nBERT 因这种训练方式而出名，擅长理解上下文。\n\nPretraining\n它常出现在预训练阶段，用海量文本先打底。\n\nSelf-supervised-learning\n它靠文本自己造题，本质上属于自监督学习。\n\nLanguage-modeling\n它是语言建模的一种变体，不走纯顺序预测路线。", "relations": { "bert": { "label": "成就…训练法", "note": "BERT 靠这种方式学上下文理解。" }, "pretraining": { "label": "属于…阶段", "note": "它常在预训练时大量使用。" }, "self-supervised-learning": { "label": "是一种…方法", "note": "不靠人工标注也能从文本里学。" }, "language-modeling": { "label": "属于…变体", "note": "它是语言建模的一种训练路线。" } } } } }, { "id": "matrix-factorization", "name": "Matrix Factorization", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "embedding" }, { "to": "unsupervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Matrix Factorization", "factExplain": "A way to split a big table into smaller tables of hidden patterns.", "humanExplain": "Imagine a school cafeteria knows your lunch mood. You skip the survey, but it still learns: tacos yes, mystery meat no.\n\nIt is used in recommendations and rating guesses. It fills blanks from a few clicks, stars, or buys.", "humanExplainDisplay": "Imagine a ==school cafeteria== knows your lunch mood.\nYou skip the survey,\nbut it still learns:\n==tacos yes, mystery meat no==.\n\nIt is used in recommendations and rating guesses.\nIt fills blanks from a few clicks,\nstars,\nor buys.", "relationsNarrative": "Embedding\nMatrix factorization often turns users and items into small vectors.\n\nUnsupervised Learning\nIt can find hidden structure when clear labels are missing.", "relations": { "embedding": { "label": "learns …", "note": "It learns vectors for users and items." }, "unsupervised-learning": { "label": "often acts as …", "note": "It can find hidden structure without clear labels." } } }, "zh": { "fullName": "矩阵分解（Matrix Factorization）", "factExplain": "把大矩阵拆成多个低维矩阵的表示方法。", "humanExplain": "相亲角里厉害的不是背简历，是一眼看出你偏爱“会聊天、爱运动、工作稳”这几条暗线。\n\n常用于推荐和评分预测，能从稀疏互动里补出你可能喜欢的内容。", "humanExplainDisplay": "相亲角里厉害的\n不是背简历，\n是一眼看出你偏爱\n==会聊天、爱运动==、\n==工作稳==这几条暗线。\n\n常用于推荐和评分预测，\n能从稀疏互动里补出\n你可能喜欢的内容。", "relationsNarrative": "Embedding\n矩阵分解的结果，常表现为用户和物品的低维向量。\n\nUnsupervised Learning\n它常在没有显式标签时，从数据里挖出隐藏结构。", "relations": { "embedding": { "label": "学出…表示", "note": "本质是在学习用户和物品向量。" }, "unsupervised-learning": { "label": "常被视作…方法", "note": "可在缺少明确标签时挖出结构。" } } } } }, { "id": "mcp", "name": "MCP", "layer": "L5", "sublayer": "architecture", "era": "2024", "publishedAt": "2026-05-23T09:55:00Z", "relations": [ { "to": "agent" }, { "to": "function-call" }, { "to": "api" }, { "to": "framework" } ], "i18n": { "en": { "fullName": "Model Context Protocol", "factExplain": "A standard way for AI to connect to outside tools and data.", "humanExplain": "MCP is like giving AI one USB-C port for all its gadgets. No more drawer full of weird chargers.\n\nAgents use it to reach files, databases, and apps more reliably. It means less custom glue code.", "humanExplainDisplay": "MCP is like giving AI\n==one USB-C port==\nfor all its gadgets.\nNo more ==drawer full of weird chargers==.\n\nAgents use it to reach files,\ndatabases,\nand apps more reliably.\nIt means less custom glue code.", "relationsNarrative": "Agent\nMCP helps an Agent reach outside tools and data more reliably.\n\nFunction-calling\nMCP gives Function-call a reusable tool connection layer.\n\nAPI\nMCP can wrap an API as a tool the model can call.\n\nFramework\nA Framework can use MCP tools inside a full workflow.", "relations": { "agent": { "label": "connects tools for …", "note": "MCP helps Agents reach outside tools and data more reliably." }, "function-call": { "label": "works with …", "note": "MCP gives Function-call a reusable way to connect tools." }, "api": { "label": "wraps … as tools", "note": "MCP can wrap an API as a tool the model can call." }, "framework": { "label": "serves as access layer for …", "note": "Frameworks can use MCP tools inside full workflows." } } }, "zh": { "fullName": "MCP", "factExplain": "一种让 AI 统一连接外部工具和数据源的协议。", "humanExplain": "它像办公室那根救命转接头，AI 想接表格、网盘、数据库，不用到处借线。\n\n常用于智能代理、编程工具和企业助手，让工具接入更统一。", "humanExplainDisplay": "它像==办公室那根救命转接头==，\nAI 想接表格、网盘、数据库，\n==不用到处借线==。\n\n常用于智能代理、编程工具和企业助手，\n让工具接入更统一。", "relationsNarrative": "Agent\nAgent 通过 MCP 更稳定地访问外部工具和数据源。\n\nFunction-calling\nMCP 为 Function-call 提供可复用的工具连接层。\n\nAPI\nAPI 通过 MCP 可被封装成模型可调用的工具。\n\nFramework\nFramework 可将 MCP 接入的工具编排进完整流程。", "relations": { "agent": { "label": "帮…接工具" }, "function-call": { "label": "与…配合" }, "api": { "label": "与…配合" }, "framework": { "label": "被…用作接入层" } } } } }, { "id": "means-ends-analysis", "name": "Means-Ends Analysis", "layer": "L2", "era": "1957", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "general-problem-solver" }, { "to": "automated-planning" }, { "to": "heuristic-search" }, { "to": "symbolic-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Means-Ends Analysis", "factExplain": "A method that compares now to the goal, then closes the gap step by step.", "humanExplain": "Means-Ends Analysis is like leaving for school in pajamas. You compare “me now” with “me ready,” then fix one gap at a time.\n\nAI uses it to plan or solve problems. It turns a big goal into smaller steps.", "humanExplainDisplay": "Means-Ends Analysis is like\nleaving for school in pajamas.\nYou compare ==me now== with ==me ready==,\nthen fix one gap at a time.\n\nAI uses it to plan or solve problems.\nIt turns a big goal into smaller steps.", "relationsNarrative": "GPS\nGPS uses Means-Ends Analysis to break a problem into doable subgoals.\n\nPlanning\nPlanning uses it to work backward from the goal to the next action.\n\nHeuristic Search\nMeans-Ends Analysis uses the current gap as a clue for the next move.\n\nSymbolic AI\nMeans-Ends Analysis came from Symbolic AI and its model of human problem solving.", "relations": { "general-problem-solver": { "label": "helps … solve", "note": "GPS uses it to break a problem into doable subgoals." }, "automated-planning": { "label": "breaks tasks for …", "note": "Planning turns the goal gap into a chain of actions." }, "heuristic-search": { "label": "guides …", "note": "Means-Ends Analysis uses the current gap to guide Heuristic Search." }, "symbolic-ai": { "label": "comes from …", "note": "Symbolic AI often broke hard problems into smaller logic steps." } } }, "zh": { "fullName": "手段-目的分析", "factExplain": "通过比较现状与目标，逐步缩小差距的求解策略。", "humanExplain": "手段-目的分析像看导航赶路：先看离终点还差多远，挑一步往哪拐，到了再看还差多少，直到归零。\n\n用于规划和解题，把目标差距拆成步骤。", "humanExplainDisplay": "手段-目的分析像\n==看导航赶路==：\n先看离终点还差多远，\n挑一步往哪拐，到了\n再看还差多少，==直到归零==。\n\n用于规划和解题，\n把目标差距拆成步骤。", "relationsNarrative": "GPS\nGPS 用手段-目的分析把问题拆成可操作子目标。\n\nPlanning\n它是自动规划中“从目标倒推行动”的早期思路。\n\nHeuristic Search\n它用当前差距作为启发，决定下一步往哪走。\n\nSymbolic AI\n它来自早期符号 AI 对人类解题过程的模拟。", "relations": { "general-problem-solver": { "label": "支撑…解题", "note": "GPS 用它把问题拆成可操作子目标。" }, "automated-planning": { "label": "帮…拆任务", "note": "规划把差距缩小成行动序列。" }, "heuristic-search": { "label": "体现…思想", "note": "用目标差距指导搜索方向。" }, "symbolic-ai": { "label": "源于…传统", "note": "早期符号 AI 偏爱分解式推理。" } } } } }, { "id": "memory-bandwidth", "name": "Memory bandwidth", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "vram" }, { "to": "inference" }, { "to": "tokens-per-second" }, { "to": "flash-attention" } ], "track": "today", "i18n": { "en": { "fullName": "Memory bandwidth", "factExplain": "The amount of data memory can send to a chip each second.", "humanExplain": "A GPU is like a race car. Memory bandwidth is the fuel pipe. If the pipe is narrow, the car cannot go fast.\n\nIt often limits inference speed. It matters a lot when reading model weights and making each new token.", "humanExplainDisplay": "A GPU is like a ==race car==.\nMemory bandwidth is the ==fuel pipe==.\nIf the pipe is narrow,\nthe car cannot go fast.\n\nIt often limits inference speed.\nIt matters a lot when reading model weights\nand making each new token.", "relationsNarrative": "VRAM\nVRAM is the storage room. Memory bandwidth decides how fast data leaves it.\n\nInference\nInference does not depend on compute alone. Memory bandwidth often sets the pace.\n\nTPS\nWhen bandwidth is tight, TPS hits a ceiling fast.\n\nFlash Attention\nFlash Attention reduces memory traffic to get around bandwidth limits.", "relations": { "vram": { "label": "sets how fast … feeds data", "note": "Big VRAM still gets stuck if bandwidth is too low." }, "inference": { "label": "limits … speed", "note": "Inference is often slowed by moving data, not math." }, "tokens-per-second": { "label": "caps …", "note": "Low bandwidth makes it hard to raise TPS." }, "flash-attention": { "label": "pushes … to cut memory traffic", "note": "Many speedups work by moving less data." } } }, "zh": { "fullName": "Memory bandwidth｜内存带宽", "factExplain": "单位时间内内存与芯片可传输的数据量。", "humanExplain": "算力再猛，也怕数据堵车；内存带宽就是那条主干道，路一堵，GPU 只能原地空踩油门。\n\n它常卡住推理速度，尤其影响参数读取和生成 token 的效率。", "humanExplainDisplay": "算力再猛，\n也怕数据==堵车==；\n内存带宽就是那条==主干道==，\n路一堵，\nGPU 只能原地空踩油门。\n\n它常卡住推理速度，\n尤其影响参数读取和生成 token 的效率。", "relationsNarrative": "VRAM\nVRAM 是数据仓库，Memory bandwidth 决定搬运有多快。\n\nInference\n推理不只看算力，常被 Memory bandwidth 卡住节奏。\n\nTokens-per-second\n带宽越吃紧，TPS 越容易碰到天花板。\n\nFlash Attention\nFlash Attention 通过减少访存压力来绕开带宽瓶颈。", "relations": { "vram": { "label": "决定…喂数速度", "note": "显存再大，带宽不够也会堵。" }, "inference": { "label": "限制…运行速度", "note": "推理常受数据搬运速度掣肘。" }, "tokens-per-second": { "label": "影响…上限", "note": "带宽不足时，生成速度很难拉高。" }, "flash-attention": { "label": "促使…优化访存", "note": "很多优化本质是在少搬数据。" } } } } }, { "id": "metric-learning", "name": "Metric Learning", "layer": "L2", "era": "1990s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "representation-learning" }, { "to": "embedding" }, { "to": "contrastive-learning" }, { "to": "k-nearest-neighbors" } ], "track": "history", "i18n": { "en": { "fullName": "Metric Learning", "factExplain": "A training method that learns how close or far examples should be.", "humanExplain": "Metric Learning is a school cafeteria seating chart for data. Dog photos sit together, and weird raccoon photos get the lonely table.\n\nYou meet it in face ID and product search. It also helps recommendations keep similar things close.", "humanExplainDisplay": "Metric Learning is a ==school cafeteria seating chart== for data.\nDog photos sit together,\nand weird raccoon photos get ==the lonely table==.\n\nYou meet it in face ID and product search.\nIt also helps recommendations keep similar things close.", "relationsNarrative": "Representation Learning\nMetric Learning is a kind of Representation Learning. It focuses on comparing examples.\n\nEmbedding\nMetric Learning turns examples into Embeddings with useful distance.\n\nContrastive Learning\nContrastive Learning often trains Metric Learning by pulling and pushing examples.\n\nKNN\nMetric Learning gives KNN a better distance to trust.", "relations": { "representation-learning": { "label": "is a kind of …", "note": "It learns a form of data that is easier to compare." }, "embedding": { "label": "learns better …", "note": "A good Embedding puts similar examples close together." }, "contrastive-learning": { "label": "often trains with …", "note": "Contrastive Learning pulls matches close and pushes non-matches away." }, "k-nearest-neighbors": { "label": "improves distance for …", "note": "KNN works better when the distance is trustworthy." } } }, "zh": { "fullName": "度量学习", "factExplain": "学习样本间距离函数的训练方法。", "humanExplain": "度量学习像武林排座次：同门师兄弟站一排，邪派卧底踢到山脚，一眼分清亲疏。\n\n用于识别、检索和推荐，让相似样本更靠近。", "humanExplainDisplay": "度量学习像武林排座次：\n==同门师兄弟==站一排，\n邪派卧底==踢到山脚==，\n一眼分清亲疏。\n\n用于识别、检索和推荐，\n让相似样本更靠近。", "relationsNarrative": "Representation Learning\n度量学习是表示学习的一种，重点学会比较样本。\n\nEmbedding\n它把样本变成向量，让远近有实际含义。\n\nContrastive Learning\n对比学习常用拉近正样本、推远负样本来训练。\n\nKNN\nKNN 依赖距离，度量学习让这个距离更靠谱。", "relations": { "representation-learning": { "label": "属于…", "note": "它学习的是更可比较的表示。" }, "embedding": { "label": "学习更好的…", "note": "好的嵌入让相似样本真的更近。" }, "contrastive-learning": { "label": "常借…训练", "note": "对比学习常用拉近拉远训练距离。" }, "k-nearest-neighbors": { "label": "改进…的距离", "note": "距离靠谱，近邻判断才靠谱。" } } } } }, { "id": "mimo-v2-5-pro-ultraspeed", "name": "MiMo UltraSpeed", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "reasoning-model" }, { "to": "tokens-per-second" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "MiMo-v2.5-Pro-UltraSpeed", "factExplain": "A reasoning model version tuned to answer fast while still staying useful.", "humanExplain": "MiMo UltraSpeed is the drive-thru lane for AI thinking. It wants your nuggets hot, right, and not next Tuesday.\n\nUse it for quick chats, writing help, and online services. The main goal is low wait time with good enough answers.", "humanExplainDisplay": "MiMo UltraSpeed is the ==drive-thru lane==\nfor AI thinking.\nIt wants your nuggets hot, right,\nand ==not next Tuesday==.\n\nUse it for quick chats,\nwriting help,\nand online services.\nThe main goal is low wait time\nwith good enough answers.", "relationsNarrative": "Reasoning-model\nMiMo UltraSpeed is a faster-leaning version of a reasoning model.\n\nTPS\nMiMo UltraSpeed usually treats output speed as a key goal.\n\nLLM\nMiMo UltraSpeed is still an LLM that can chat and write.", "relations": { "reasoning-model": { "label": "is a fast version of …", "note": "It is a reasoning model with a strong speed focus." }, "tokens-per-second": { "label": "aims for higher …", "note": "This version usually treats output speed as a key goal." }, "llm": { "label": "is a kind of …", "note": "It is still a large model that can chat and write." } } }, "zh": { "fullName": "MiMo-v2.5-Pro-UltraSpeed 模型", "factExplain": "一个强调速度与性能平衡的推理模型版本。", "humanExplain": "像写字楼电梯里的急行梯：不跟你慢慢晃，按钮一按就直奔楼层，先把人尽快送到位。\n\n适合低延迟对话、写作和在线服务，重点是在够用前提下先快起来。", "humanExplainDisplay": "像写字楼电梯里的==急行梯==：\n不跟你慢慢晃，\n按钮一按就==直奔楼层==，\n先把人尽快送到位。\n\n适合低延迟对话、\n写作和在线服务，\n重点是在够用前提下先快起来。", "relationsNarrative": "Reasoning-model\n它可看作更偏速度取向的推理模型版本。\n\nTokens-per-second\n这类版本通常会把生成速度当成重点指标。\n\nLLM\n它本质上仍属于可对话、可生成内容的大模型。", "relations": { "reasoning-model": { "label": "可视作…变体", "note": "它更像强调速度取向的推理模型。" }, "tokens-per-second": { "label": "追求更高…", "note": "这类版本通常会优先优化生成速度。" }, "llm": { "label": "属于…一类", "note": "本质上仍是可对话生成的大模型。" } } } } }, { "id": "minimax-m3", "name": "MiniMax M3", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "chatgpt" }, { "to": "maas-model-as-a-service" } ], "track": "today", "i18n": { "en": { "fullName": "MiniMax M3 Model", "factExplain": "A general-purpose large AI model from MiniMax.", "humanExplain": "MiniMax M3 is the office pinch-hitter. If the slides catch fire, it grabs the clicker and covers the shift.\n\nIt often works as the base for general AI assistants. You may meet it in chat tools or company office apps.", "humanExplainDisplay": "MiniMax M3 is the ==office pinch-hitter==.\nIf the slides catch fire,\nit grabs the clicker\nand ==covers the shift==.\n\nIt often works as the base\nfor general AI assistants.\nYou may meet it in chat tools\nor company office apps.", "relationsNarrative": "LLM\nMiniMax M3 is a kind of large language model.\n\nChatGPT\nMiniMax M3 is often compared with ChatGPT for general assistant use.\n\nMaaS\nMiniMax M3 can be offered as a cloud model service.", "relations": { "llm": { "label": "is a kind of …", "note": "MiniMax M3 is a general-purpose LLM." }, "chatgpt": { "label": "is often compared with …", "note": "Both are built for general chat and assistant tasks." }, "maas-model-as-a-service": { "label": "can be offered through …", "note": "Models like this are often delivered as cloud services." } } }, "zh": { "fullName": "MiniMax M3 模型", "factExplain": "MiniMax 推出的通用大模型。", "humanExplain": "放职场里，它像那种临时被抓去救场的人：PPT能改、会能开、客户也能接，主打一个==哪都能顶班==。\n\n常做通用助手底座，用在聊天、办公和企业应用。", "humanExplainDisplay": "放职场里，\n它像那种临时被抓去救场的人：\nPPT能改、会能开、\n客户也能接，主打一个==哪都能顶班==。\n\n常做通用助手底座，\n用在聊天、办公\n和企业应用。", "relationsNarrative": "LLM\n它本质上属于大语言模型这一类。\n\nChatGPT\n它常在通用助手体验上被拿来与 ChatGPT 对比。\n\nMaaS\n这类模型通常会以云端模型服务的方式提供。", "relations": { "llm": { "label": "属于…一类", "note": "它本质上是一种通用大语言模型。" }, "chatgpt": { "label": "常被拿来对比", "note": "都面向通用对话与助手场景。" }, "maas-model-as-a-service": { "label": "可通过…提供", "note": "这类模型常以云服务方式交付。" } } } } }, { "id": "minimax-search", "name": "Minimax Search", "layer": "L2", "era": "1950", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "alpha-beta-pruning" }, { "to": "monte-carlo-tree-search" }, { "to": "automated-planning" }, { "to": "a-search" } ], "track": "history", "i18n": { "en": { "fullName": "Minimax Search", "factExplain": "Search for the best move against the opponent’s best reply.", "humanExplain": "Minimax is chess with your annoyingly smart cousin. Your move only counts after their nastiest comeback.\n\nIt looks ahead through turns, then picks the safest strong move. You meet it in board games and planning with opponents.", "humanExplainDisplay": "Minimax is ==chess with your annoyingly smart cousin==.\nYour move only counts\nafter their ==nastiest comeback==.\n\nIt looks ahead through turns,\nthen picks the safest strong move.\nYou meet it in board games\nand planning with opponents.", "relationsNarrative": "Α-β Pruning\nΑ-β Pruning helps Minimax search faster by skipping useless branches.\n\nMCTS\nMinimax and MCTS both choose game moves, but they explore in different ways.\n\nPlanning\nMinimax is a classic planning method for games with an opponent.\n\nA* Search\nMinimax and A* Search both search paths, but A* Search does not plan against an opponent.", "relations": { "alpha-beta-pruning": { "label": "uses … to search faster", "note": "Α-β Pruning cuts useless branches without changing the result." }, "monte-carlo-tree-search": { "label": "chooses game moves like …", "note": "Both choose game moves, but they search in different ways." }, "automated-planning": { "label": "is a classic … method", "note": "Minimax is a classic planning method for settings with an opponent." }, "a-search": { "label": "searches paths like …", "note": "Both search for next moves, but A* Search does not model an opponent." } } }, "zh": { "fullName": "极大极小搜索", "factExplain": "在对抗场景中按双方最优反应来选行动的搜索方法。", "humanExplain": "武侠过招不是只想自己怎么打，还得先替对面把后手演完，活下来那招才算数。\n\n它常用于棋类博弈和对抗规划，帮系统选更稳的下一步。", "humanExplainDisplay": "武侠过招不是只想\n自己怎么打，\n还得先替对面把==后手演完==，\n活下来那招才==算数==。\n\n它常用于棋类博弈\n和对抗规划，\n帮系统选更稳的下一步。", "relationsNarrative": "Alpha-Beta Pruning\n它常配合极大极小搜索，用剪枝减少要看的分支。\n\nMonte-Carlo Tree Search\n两者都用于博弈决策，但一个偏穷举评估，一个偏采样探索。\n\nAutomated Planning\n它是经典搜索式规划方法，尤其适合双方对抗场景。\n\nA* Search\n两者都在搜索行动路径，但 A* 不专门处理对手博弈。", "relations": { "alpha-beta-pruning": { "label": "用…加速搜索", "note": "它在不改结果前提下剪掉无用分支。" }, "monte-carlo-tree-search": { "label": "与…同做博弈", "note": "两者都用于游戏决策，但思路不同。" }, "automated-planning": { "label": "属于…经典方法", "note": "它是对抗环境里的经典规划套路。" }, "a-search": { "label": "和…同属搜索", "note": "都在找下一步，但目标函数不同。" } } } } }, { "id": "minimum-description-length", "name": "MDL", "layer": "L2", "era": "1978", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "information-theory" }, { "to": "regularization" }, { "to": "bias-variance-tradeoff" }, { "to": "statistical-learning-theory" } ], "track": "history", "i18n": { "en": { "fullName": "Minimum Description Length", "factExplain": "A rule that picks the model with the shortest total code.", "humanExplain": "MDL is like explaining a broken cookie jar to your mom. Mom counts the story words plus the excuse words. The shortest total wins her trust.\n\nIt helps choose between models. It adds model size plus mistakes, so no model memorizes noise.", "humanExplainDisplay": "MDL is like explaining a ==broken cookie jar== to your mom.\nMom counts the ==story words==\nplus the ==excuse words==.\nThe shortest total wins her trust.\n\nIt helps choose between models.\nIt adds model size plus mistakes,\nso no model memorizes noise.", "relationsNarrative": "Information Theory\nMDL uses code length to measure both model size and mistakes.\n\nRegularization\nRegularization often puts the short-explanation idea into the training score.\n\nBias-Variance Tradeoff\nMDL warns the model not to memorize noise just to lower errors.\n\nSLT\nMDL is a model choice rule that prefers simple explanations.", "relations": { "information-theory": { "label": "measures simplicity with …", "note": "Code length turns complexity into a number." }, "regularization": { "label": "penalizes complex models like …", "note": "Both stop a model from treating noise as a rule." }, "bias-variance-tradeoff": { "label": "balances complexity in …", "note": "Too short can miss patterns. Too long can memorize noise." }, "statistical-learning-theory": { "label": "acts as a choice rule in …", "note": "It gives model choice an information-based view." } } }, "zh": { "fullName": "最小描述长度", "factExplain": "用最短总编码选择模型的原则。", "humanExplain": "最小描述长度像记外语语法：规则要背，例外也要背，两本账加起来最薄的才是好规则。\n\n用于模型选择，也防过拟合：模型加误差，一起算总长。", "humanExplainDisplay": "最小描述长度像记外语语法：\n==规则要背==，\n==例外也要背==，\n两本账加起来最薄的，\n才是好规则。\n\n用于模型选择，也防过拟合：\n模型加误差，一起算总长。", "relationsNarrative": "Information Theory\nMDL 用编码长度统一模型复杂度和拟合误差。\n\nRegularization\n正则化常把“短解释”思想写进损失函数。\n\nBias-Variance Tradeoff\n它提醒模型别为降低误差，把噪声也背下来。\n\nStatistical Learning Theory\n它是模型选择中偏好简单解释的一种原则。", "relations": { "information-theory": { "label": "借…衡量简洁性", "note": "编码长度把复杂度变成数字。" }, "regularization": { "label": "像…惩罚复杂模型", "note": "两者都防模型把噪声当规律。" }, "bias-variance-tradeoff": { "label": "权衡…中的复杂度", "note": "太短易欠拟合，太长易过拟合。" }, "statistical-learning-theory": { "label": "属于…的选择原则", "note": "它给模型选择提供信息论视角。" } } } } }, { "id": "mixture-of-experts-moe", "name": "MoE", "aliases": [ "混合专家" ], "layer": "L3", "era": "2017", "publishedAt": "2026-05-30T03:10:23.224Z", "relations": [ { "to": "llm" }, { "to": "parameter" }, { "to": "gpu" }, { "to": "inference" } ], "track": "today", "i18n": { "en": { "fullName": "Mixture of Experts", "factExplain": "A model design that turns on only some parameters for each input.", "humanExplain": "MoE is like a giant diner with many cooks. Order pancakes, and only the breakfast cooks jump in. The sushi chef stays seated.\n\nIt helps build huge neural networks without waking every parameter each time. You meet it in some LLMs during training and when they run.", "humanExplainDisplay": "MoE is like a giant diner\nwith many cooks.\nOrder pancakes,\nand only the ==breakfast cooks== jump in.\nThe ==sushi chef== stays seated.\n\nIt helps build huge neural networks\nwithout waking every parameter each time.\nYou meet it in some LLMs\nduring training and when they run.", "relationsNarrative": "LLM\nMoE can be an LLM architecture for bigger models with better cost control.\n\nParameter\nMoE may have many parameters, but one input turns on only some of them.\n\nGPU\nMoE tries to use GPU compute only on the most needed parts.\n\nInference\nMoE changes compute during inference, so cost and speed can change.", "relations": { "llm": { "label": "can be an … architecture", "note": "Some LLMs use MoE to grow with less wasted compute." }, "parameter": { "label": "turns on only some …", "note": "MoE has many parameters, but each input uses only some." }, "gpu": { "label": "helps save … compute", "note": "MoE tries to spend GPU work only on the needed parts." }, "inference": { "label": "changes … cost", "note": "MoE changes how much work happens during inference." } } }, "zh": { "fullName": "专家混合模型", "factExplain": "一种处理输入时只激活部分参数的模型结构。", "humanExplain": "它像大厂分工：需求一来，不全员开会，只叫对口专家加班。\n\n常用于大模型扩容，省算力，但路由和部署更难。", "humanExplainDisplay": "它像==大厂分工==：\n需求一来，\n不全员开会，\n只叫==对口专家加班==。\n\n常用于大模型扩容，\n省算力，\n但路由和部署更难。", "relationsNarrative": "LLM\nMoE 是不少 LLM 会采用的模型架构，用来在更大规模下兼顾成本与能力。\n\nParameter\nMoE 往往拥有很多参数，但处理一次输入时只会激活其中一部分。\n\nGPU\nMoE 的思路之一，是把算力花在更需要的部分，减少每次都全模型出动。\n\nInference\nMoE 会影响推理时的计算分配，因此也会影响部署成本与速度。", "relations": { "llm": { "label": "可作为…架构", "note": "不少 LLM 会用 MoE 提升规模效率。" }, "parameter": { "label": "只激活部分…", "note": "MoE 参数很多，但每次只用一部分。" }, "gpu": { "label": "帮助省…算力", "note": "它想用更少计算撑起更大模型。" }, "inference": { "label": "影响…成本", "note": "MoE 会改变模型推理时的计算方式。" } } } } }, { "id": "mlp", "name": "MLP", "layer": "L3", "era": "1986", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "backpropagation" }, { "to": "transformer" }, { "to": "parameter" } ], "track": "history", "i18n": { "en": { "fullName": "Multilayer Perceptron", "factExplain": "A neural network with several fully connected layers, moving data one way.", "humanExplain": "An MLP is like a row of picky lunch judges. Each judge tastes the whole soup, then passes a score down the line.\n\nIt is a basic neural-network shape. It often sorts things into groups. It also sits as a small block inside big AI models.", "humanExplainDisplay": "An MLP is like a row of ==picky lunch judges==.\nEach judge tastes the ==whole soup==,\nthen passes a score down the line.\n\nIt is a basic neural-network shape.\nIt often sorts things into groups.\nIt also sits as a small block inside big AI models.", "relationsNarrative": "Neural-network\nAn MLP is one of the classic basic shapes in the neural-network family.\n\nBackpropagation\nAn MLP usually uses Backprop to adjust weights in each layer.\n\nTransformer\nIn a Transformer, an MLP often works as the feed-forward network block.\n\nParameter\nAn MLP has weights and biases, and both are learned Parameters.", "relations": { "neural-network": { "label": "is a kind of …", "note": "It is one of the classic neural-network shapes." }, "backpropagation": { "label": "learns with …", "note": "Backprop usually updates the weights in each layer." }, "transformer": { "label": "works inside …", "note": "In a Transformer, an MLP often works as the feed-forward block." }, "parameter": { "label": "is made of …", "note": "Its weights and biases are learned parameters." } } }, "zh": { "fullName": "多层感知机（Multilayer Perceptron）", "factExplain": "由多层全连接层组成的前馈神经网络。", "humanExplain": "它像相亲时层层把关：先看照片简历，再聊三观条件，最后才决定要不要继续。\n\n是神经网络的基础结构，常做分类预测，也常当大模型里的小模块。", "humanExplainDisplay": "它像相亲时==层层把关==：\n先看照片简历，\n再聊三观条件，\n最后才==决定要不要继续==。\n\n是神经网络的基础结构，\n常做分类预测，\n也常当大模型里的小模块。", "relationsNarrative": "Neural-network\n它是神经网络家族里最经典的基础结构之一。\n\nBackpropagation\n它通常靠反向传播来调整各层权重。\n\nTransformer\n在 Transformer 里，它常充当前馈网络模块。\n\nParameter\n它的连接权重和偏置，都是需要学习的参数。", "relations": { "neural-network": { "label": "属于…的一种", "note": "它是经典神经网络结构之一。" }, "backpropagation": { "label": "靠…学参数", "note": "通常用反向传播更新各层权重。" }, "transformer": { "label": "在…里当模块", "note": "它常作为其中的前馈子层。" }, "parameter": { "label": "由…组成", "note": "它的权重和偏置都是参数。" } } } } }, { "id": "mmlu", "name": "MMLU", "layer": "L4", "era": "2020", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "model-leaderboard" }, { "to": "benchmark-contamination" }, { "to": "question-answering" } ], "track": "history", "i18n": { "en": { "fullName": "Massive Multitask Language Understanding", "factExplain": "A benchmark for testing language model knowledge with multiple-choice questions across many subjects.", "humanExplain": "MMLU is one huge exam day for AI. Math sits by law. Medicine brings a clipboard.\n\nIt appears on leaderboards to check school knowledge and expert knowledge. A high score does not mean real job skills.", "humanExplainDisplay": "MMLU is one ==huge exam day== for AI.\nMath sits by law.\n==Medicine brings a clipboard.==\n\nIt appears on leaderboards\nto check school knowledge and expert knowledge.\nA high score does not mean real job skills.", "relationsNarrative": "LLM\nMMLU measures an LLM’s knowledge across many subjects.\n\nLeaderboard\nLeaderboards often use MMLU scores to compare models.\n\nBenchmark contamination\nIf test questions leak into training, MMLU scores can look too high.\n\nQA\nMMLU mostly uses multiple-choice QA questions.", "relations": { "llm": { "label": "tests … knowledge", "note": "MMLU measures an LLM’s knowledge with a question set." }, "model-leaderboard": { "label": "helps rank …", "note": "Leaderboards often compare models with MMLU scores." }, "benchmark-contamination": { "label": "can suffer from …", "note": "Leaked questions can make MMLU scores look too high." }, "question-answering": { "label": "uses … format", "note": "Most MMLU tasks are multiple-choice QA questions." } } }, "zh": { "fullName": "Massive Multitask Language Understanding，大规模多任务语言理解", "factExplain": "用多学科选择题评测语言模型知识的基准。", "humanExplain": "MMLU 像给 AI 办超级联考：文理医法全上桌，看是真学霸还是会蒙。\n\n常用于模型排行榜，衡量通识与专业知识，但分数不等于会干活。", "humanExplainDisplay": "MMLU 像给 AI 办\n==超级联考==：\n文理医法全上桌，\n看是真学霸还是==会蒙==。\n\n常用于模型排行榜，\n衡量通识与专业知识，\n但分数不等于会干活。", "relationsNarrative": "LLM\nMMLU 常用来衡量 LLM 的多学科知识水平。\n\nLeaderboard\n排行榜常用 MMLU 分数比较模型强弱。\n\nBenchmark Contamination\n若题目进了训练集，MMLU 分数会虚高。\n\nQA\nMMLU 的题目主要以选择问答形式呈现。", "relations": { "llm": { "label": "评测…知识", "note": "MMLU 用题库衡量 LLM 知识水平。" }, "model-leaderboard": { "label": "支撑…排名", "note": "排行榜常用 MMLU 分数比较模型。" }, "benchmark-contamination": { "label": "容易遭遇…", "note": "题目泄漏会让分数虚高。" }, "question-answering": { "label": "采用…形式", "note": "多数学科题以选择问答呈现。" } } } } }, { "id": "mnist", "name": "MNIST", "layer": "L4", "era": "1998", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "image-classification" }, { "to": "cnn" }, { "to": "supervised-learning" }, { "to": "computer-vision" } ], "track": "history", "i18n": { "en": { "fullName": "Handwritten Digit Dataset", "factExplain": "A classic image dataset for teaching AI to read handwritten digits.", "humanExplain": "MNIST is AI’s kindergarten number sheet. It reads wobbly 0s to 9s, so do not hand it the SAT yet.\n\nPeople use it to practice image sorting and test new ideas. It is so simple, it feels like an AI eye exam.", "humanExplainDisplay": "MNIST is AI’s ==kindergarten number sheet==.\nIt reads ==wobbly 0s to 9s==,\nso do not hand it the SAT yet.\n\nPeople use it to practice image sorting\nand test new ideas.\nIt is so simple,\nit feels like an AI eye exam.", "relationsNarrative": "Image Class.\nMNIST is the classic starter exercise for Image Class.\n\nCNN\nEarly CNNs often used MNIST to test handwritten digit reading.\n\nSupervised Learning\nMNIST has a label for each digit image, so it fits Supervised Learning.\n\nComputer Vision\nMNIST is a common old dataset for learning Computer Vision.", "relations": { "image-classification": { "label": "used to practice …", "note": "MNIST is a classic first task for Image Class." }, "cnn": { "label": "often tests …", "note": "Early CNNs often used MNIST to test digit reading." }, "supervised-learning": { "label": "provides samples for …", "note": "Each digit image has the correct label." }, "computer-vision": { "label": "is beginner data for …", "note": "MNIST is an old practice book for Computer Vision." } } }, "zh": { "fullName": "手写数字数据集", "factExplain": "用于手写数字识别的经典图像数据集。", "humanExplain": "MNIST 是 AI 的描红本：歪歪扭扭的 0 到 9，认不清别急着考清华。\n\n用于练图像分类和测算法，简单到更像入门体检。", "humanExplainDisplay": "MNIST 是 AI 的\n==描红本==：\n歪歪扭扭的 ==0 到 9==，\n认不清别急着考清华。\n\n用于练图像分类和测算法，\n简单到，\n更像入门体检。", "relationsNarrative": "Image-classification\nMNIST 是图像分类里最经典的入门练习题。\n\nCNN\n早期 CNN 常用 MNIST 验证识别手写数字的效果。\n\nSupervised Learning\nMNIST 每张数字图都有标签，适合监督学习训练。\n\nComputer Vision\nMNIST 是计算机视觉学习中常见的老牌数据集。", "relations": { "image-classification": { "label": "用于练习…", "note": "MNIST 是图像分类的经典入门题。" }, "cnn": { "label": "常用来测试…", "note": "早期 CNN 常拿 MNIST 验证效果。" }, "supervised-learning": { "label": "提供…样本", "note": "每张数字图都有对应标签。" }, "computer-vision": { "label": "属于…入门数据", "note": "它是视觉任务的老牌练习册。" } } } } }, { "id": "model-abliteration", "name": "Abliteration", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "alignment" }, { "to": "fine-tuning" }, { "to": "open-weights" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Model abliteration", "factExplain": "A targeted change that weakens a model’s specific skill or habit.", "humanExplain": "Model abliteration is not smashing the vending machine. It is rewiring one sketchy button, so it stops selling chaos soda.\n\nPeople use it for safety work and model editing. It can also weaken useful skills by accident.", "humanExplainDisplay": "Model abliteration is not smashing the ==vending machine==.\nIt is rewiring one ==sketchy button==,\nso it stops selling chaos soda.\n\nPeople use it for safety work\nand model editing.\nIt can also weaken useful skills\nby accident.", "relationsNarrative": "Alignment\nAbliteration can help Alignment by pushing down risky behavior.\n\nFine-tuning\nAbliteration is targeted editing, not normal extra training.\n\nOpen weights\nOpen weights make this kind of direct model edit easier.\n\nLLM\nAbliteration changes some skills or habits inside an LLM.", "relations": { "alignment": { "label": "can serve as …", "note": "It can push down model behavior people do not want." }, "fine-tuning": { "label": "differs from …", "note": "It is more like targeted editing than more training." }, "open-weights": { "label": "is easier with …", "note": "Open weights let outsiders make direct model edits." }, "llm": { "label": "edits inside an …", "note": "It changes some skills or habits inside the LLM." } } }, "zh": { "fullName": "Model abliteration（模型去特性/抹除）", "factExplain": "用定向修改削弱模型的特定能力或倾向。", "humanExplain": "像把小区门卫的“不能进”口头禅抹掉，谁来都笑脸放行。\n\n常用于制作少拒答模型，也会放大越狱和滥用风险。", "humanExplainDisplay": "像把小区门卫的\n==“不能进”口头禅抹掉==，\n谁来都==笑脸放行==。\n\n常用于制作少拒答模型，\n也会放大越狱和滥用风险。", "relationsNarrative": "Alignment\nAbliteration 可作为对齐手段，压制特定危险行为。\n\nFine-tuning\n它偏向定向删改，不等于常规的继续训练。\n\nOpen-weights\n开放权重让研究者更容易做这类定向改造。\n\nLLM\nAbliteration 直接作用于 LLM 的部分能力或倾向。", "relations": { "alignment": { "label": "作为…手段", "note": "它可用于压制不想要的模型行为。" }, "fine-tuning": { "label": "区别于…训练", "note": "它更像定向删改，不只是继续训练。" }, "open-weights": { "label": "常用于…模型", "note": "开放权重让外部更容易做定向改造。" }, "llm": { "label": "作用于…内部", "note": "它直接改动大模型的部分行为倾向。" } } } } }, { "id": "model-based-reinforcement-learning", "name": "MBRL", "layer": "L2", "era": "1990", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "world-model" }, { "to": "markov-decision-process" }, { "to": "monte-carlo-tree-search" } ], "track": "history", "i18n": { "en": { "fullName": "Model-Based Reinforcement Learning", "factExplain": "A reinforcement learning method using a learned world model to choose actions.", "humanExplain": "MBRL is like trying a skateboard trick in a video game first. Your knees send a thank-you card.\n\nIt learns a mini world, then tries choices there before acting. Robots and self-driving cars use it to cut real test runs, but a bad model can steer them wrong.", "humanExplainDisplay": "MBRL is like trying a ==skateboard trick==\nin a ==video game== first.\nYour knees send a thank-you card.\n\nIt learns a mini world,\nthen tries choices there before acting.\nRobots and self-driving cars use it\nto cut real test runs,\nbut a bad model can steer them wrong.", "relationsNarrative": "RL\nMBRL belongs to RL and learns the environment first.\n\nWorld model\nA World model acts like its practice world for predictions.\n\nMDP\nMBRL often uses an MDP to plan from state changes.\n\nMCTS\nMCTS can search a few moves inside its practice world.", "relations": { "reinforcement-learning": { "label": "is a branch of …", "note": "MBRL is a core branch of RL." }, "world-model": { "label": "builds with …", "note": "A World model predicts what may happen next." }, "markov-decision-process": { "label": "plans in …", "note": "An MDP frames choices as states, actions, and changes." }, "monte-carlo-tree-search": { "label": "searches with …", "note": "MCTS can test moves inside the learned model." } } }, "zh": { "fullName": "基于模型的强化学习（Model-Based Reinforcement Learning）", "factExplain": "先学习环境模型，再用它辅助决策的强化学习方法。", "humanExplain": "武侠里真高手不会回回拿命试招，先在脑子里把对手路数过几遍，再决定这一剑出不出。\n\n常用于机器人和自动驾驶，能减少真实试错；但内部模型不准时会带偏决策。", "humanExplainDisplay": "武侠里真高手\n不会回回\n==拿命试招==，\n先在脑子里把\n对手路数过几遍，\n再决定这一剑\n==出不出==。\n\n常用于机器人和自动驾驶，\n能减少真实试错；\n但内部模型不准时，\n会带偏决策。", "relationsNarrative": "Reinforcement-learning\n它是强化学习的重要分支，区别在于先学环境再决策。\n\nWorld-model\n世界模型常充当内部模拟器，帮助它预测后果。\n\nMarkov-decision-process\n它通常建立在状态、动作与转移的决策框架上。\n\nMonte-carlo-tree-search\n它可结合搜索，在模拟环境里先试走几步。", "relations": { "reinforcement-learning": { "label": "属于…分支", "note": "它是强化学习的一条核心路线。" }, "world-model": { "label": "常用…建环境", "note": "世界模型常被拿来预测环境变化。" }, "markov-decision-process": { "label": "在…里规划", "note": "它通常基于状态转移做决策。" }, "monte-carlo-tree-search": { "label": "配合…试走", "note": "可先在内部模拟里搜索动作。" } } } } }, { "id": "model-calibration", "name": "Model Calibration", "layer": "L2", "era": "2000s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "model-uncertainty" }, { "to": "ai-abstention" }, { "to": "softmax" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Model Calibration", "factExplain": "A technique that makes model confidence match its real chance of being right.", "humanExplain": "Model calibration is like a weather app with manners. If it says 70% chance of rain, it should rain about 7 out of 10 times.\n\nYou meet it in hospitals, banks, and self-driving cars. It helps AI know when to call a human.", "humanExplainDisplay": "Model calibration is like a weather app with manners.\nIf it says ==70% chance of rain==,\nit should rain about ==7 out of 10== times.\n\nYou meet it in hospitals, banks, and self-driving cars.\nIt helps AI know when to call a human.", "relationsNarrative": "Model uncertainty\nModel calibration makes uncertainty scores closer to real risk.\n\nAI Abstention\nAI can refuse to answer when its confidence is trustworthy.\n\nSoftmax\nSoftmax scores can look like probabilities, and calibration helps fix them.\n\nClassification\nClassification often needs calibrated class probabilities.", "relations": { "model-uncertainty": { "label": "fixes … readings", "note": "Calibration makes uncertainty scores easier to trust." }, "ai-abstention": { "label": "supports … choices", "note": "The AI can stay quiet only when confidence means something." }, "softmax": { "label": "corrects … scores", "note": "Softmax scores can look like probabilities but be wrong." }, "classification": { "label": "improves … probabilities", "note": "Classifiers need class confidence scores people can trust." } } }, "zh": { "fullName": "模型校准", "factExplain": "使模型置信度接近真实正确率的技术。", "humanExplain": "模型校准像相亲照去滤镜：说七分就真七分，别把普通自信拍成天仙下凡。\n\n用于医疗、风控、自动驾驶，判断何时交给人。", "humanExplainDisplay": "模型校准像==相亲照去滤镜==：\n==说七分就真七分==，\n别把普通自信\n拍成天仙下凡。\n\n用于医疗、风控、自动驾驶，\n判断何时，\n交给人。", "relationsNarrative": "Model uncertainty\n模型校准让不确定性分数更接近真实风险。\n\nAI Abstention\n置信度可信时，系统才知道该不该拒答。\n\nSoftmax\nSoftmax 分数常被当概率，校准可纠偏。\n\nClassification\n分类任务常需要校准后的类别概率。", "relations": { "model-uncertainty": { "label": "校正…读数", "note": "校准让不确定性分数更可信。" }, "ai-abstention": { "label": "支撑…决策", "note": "置信度可信，才敢选择不回答。" }, "softmax": { "label": "修正…分数", "note": "Softmax 分数常像概率却不准。" }, "classification": { "label": "改善…概率", "note": "分类器需要可靠的类别置信度。" } } } } }, { "id": "model-compression", "name": "Model Compression", "layer": "L2", "era": "2006", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "quantization" }, { "to": "distillation" }, { "to": "local-llm" }, { "to": "vram" } ], "track": "history", "i18n": { "en": { "fullName": "Model Compression", "factExplain": "Making a model smaller and lighter while trying to keep its skill.", "humanExplain": "Model compression is like repacking a huge camping backpack. Keep the snacks. Leave the cast-iron pan.\n\nYou meet it when AI must run on a laptop or phone. It uses less VRAM and often runs faster.", "humanExplainDisplay": "Model compression is like repacking\na ==huge camping backpack==.\nKeep the snacks.\nLeave the ==cast-iron pan==.\n\nYou meet it when AI must run\non a laptop or phone.\nIt uses less VRAM\nand often runs faster.", "relationsNarrative": "Quantization\nQuantization is one of the most common ways to compress a model.\n\nDistillation\nDistillation helps move a big model’s skill into a smaller model.\n\nLocal-LLM\nModel compression helps a Local-LLM run on a normal device.\n\nVRAM\nSmaller models usually use less VRAM while running.", "relations": { "quantization": { "label": "often uses …", "note": "Quantization is one of the most common ways to compress a model." }, "distillation": { "label": "can shrink via …", "note": "Distillation moves big-model skill into a smaller model." }, "local-llm": { "label": "helps … run on devices", "note": "Compression makes local models easier to run on your own device." }, "vram": { "label": "uses less …", "note": "A smaller model usually needs less VRAM while running." } } }, "zh": { "fullName": "Model Compression｜模型压缩", "factExplain": "在尽量少掉效果的前提下，把模型变小变轻。", "humanExplain": "模型压缩像搬家塞行李箱：衣服卷一卷、盒子拆一拆，箱子小了，东西还尽量别落下。\n\n常用于端侧和本地部署，让模型更省显存、跑得更快。", "humanExplainDisplay": "模型压缩像搬家塞行李箱：\n衣服==卷一卷==、\n盒子拆一拆，\n箱子==小了==，东西还尽量别落下。\n\n常用于端侧和本地部署，\n让模型更省显存、\n跑得更快。", "relationsNarrative": "Quantization\n量化是模型压缩最常见、最实用的做法之一。\n\nDistillation\n蒸馏常用来把大模型能力装进更小模型里。\n\nLocal-LLM\n模型压缩能让本地模型更容易在设备上跑起来。\n\nVRAM\n模型越小，运行时通常占用的显存也越少。", "relations": { "quantization": { "label": "常靠…落地", "note": "量化是最常见的压缩手段之一。" }, "distillation": { "label": "可用…瘦身", "note": "蒸馏把大模型能力迁到小模型。" }, "local-llm": { "label": "帮助…上设备", "note": "压缩后更适合本地设备运行。" }, "vram": { "label": "减少…占用", "note": "模型变小后通常更省显存。" } } } } }, { "id": "model-free-reinforcement-learning", "name": "Model-free-reinforcement-learning", "layer": "L2", "era": "1989", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "model-based-reinforcement-learning" }, { "to": "q-learning" }, { "to": "policy-gradient" }, { "to": "deep-reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Model-Free Reinforcement Learning", "factExplain": "AI learns what to do by trying, without first building a world model.", "humanExplain": "It is like learning a video game boss fight with no guide. You get squished, respawn, and slowly stop doing dumb stuff.\n\nThe AI tries actions and keeps the ones that pay off. You see it in game bots and robot control.", "humanExplainDisplay": "It is like learning a video game boss fight\nwith ==no guide==.\nYou ==get squished==, respawn,\nand slowly stop doing dumb stuff.\n\nThe AI tries actions\nand keeps the ones that pay off.\nYou see it in game bots\nand robot control.", "relationsNarrative": "MBRL\nModel-free learning skips the world model and learns by trial and error.\n\nQ-Learning\nQ-Learning is a classic model-free method for learning action scores.\n\nPolicy Gradient\nPolicy Gradient is another model-free path that learns the policy directly.\n\nDeep RL\nDeep RL often uses neural networks to run model-free methods.", "relations": { "model-based-reinforcement-learning": { "label": "contrasts with …", "note": "Model-free learns by trial and error. MBRL learns a world model first." }, "q-learning": { "label": "includes …", "note": "Q-Learning is a classic model-free way to score actions." }, "policy-gradient": { "label": "includes …", "note": "Policy Gradient directly learns the action rule." }, "deep-reinforcement-learning": { "label": "often combines with …", "note": "Deep RL often uses neural networks for model-free learning." } } }, "zh": { "fullName": "Model-Free Reinforcement Learning／无模型强化学习", "factExplain": "不显式学习环境模型，直接从交互中学策略或价值。", "humanExplain": "这路子像新手跑业务不背话术，先一单单碰壁，慢慢知道哪句会冷场、哪种客户更容易签。\n\n常用于游戏对战和机器人控制，靠反复试错学会决策。", "humanExplainDisplay": "这路子像新手跑业务\n不背==话术==，\n先一单单碰壁，\n慢慢知道哪句会冷场、\n哪种客户更容易\n==签==。\n\n常用于游戏对战\n和机器人控制，\n靠反复试错\n学会决策。", "relationsNarrative": "Model-Based Reinforcement Learning\n它不先学环境模型，而是直接靠交互试错学习。\n\nQ-Learning\nQ-Learning 是无模型强化学习的经典价值学习方法。\n\nPolicy Gradient\n策略梯度是无模型强化学习的另一条直接学策略路线。\n\nDeep RL\n深度强化学习常用神经网络实现无模型方法。", "relations": { "model-based-reinforcement-learning": { "label": "对比…路线", "note": "一个靠试错，一个先学环境模型。" }, "q-learning": { "label": "包含…方法", "note": "Q-Learning 是它的经典代表。" }, "policy-gradient": { "label": "包含…方法", "note": "策略梯度也是常见做法。" }, "deep-reinforcement-learning": { "label": "常与…结合", "note": "配合神经网络后能力更强。" } } } } }, { "id": "model-leaderboard", "name": "Leaderboard", "layer": "L4", "era": "2023", "publishedAt": "2026-05-30T03:10:23.229Z", "relations": [ { "to": "benchmark-contamination" }, { "to": "third-party-ai-evaluation" }, { "to": "llm-as-a-judge" }, { "to": "frontier-model" } ], "track": "today", "i18n": { "en": { "fullName": "Model Leaderboard", "factExplain": "A ranked list of AI models based on the same tests.", "humanExplain": "A model leaderboard is class test results pinned in the hallway. The top score gets stared at like it brought pizza.\n\nPeople use it to pick models and spot trends. But a high score does not always mean it works best for you.", "humanExplainDisplay": "A model leaderboard is ==class test results==\npinned in the hallway.\nThe ==top score== gets stared at\nlike it brought pizza.\n\nPeople use it to pick models\nand spot trends.\nBut a high score does not always mean\nit works best for you.", "relationsNarrative": "Benchmark contamination\nLeaked tests can make leaderboard scores look fake.\n\nThird-party AI evaluation\nThird-party eval adds another view beyond one score.\n\nLLM-as-a-judge\nSome leaderboards use LLM-as-a-judge to help grade answers.\n\nFrontier model\nFrontier models often race each other on leaderboards.", "relations": { "benchmark-contamination": { "label": "can be inflated by …", "note": "Leaked test questions can make leaderboard scores look too high." }, "third-party-ai-evaluation": { "label": "often checks against …", "note": "Third-party eval adds another view beyond one score." }, "llm-as-a-judge": { "label": "may use … to score", "note": "Some leaderboards let models help grade answers." }, "frontier-model": { "label": "often ranks …", "note": "Frontier models often fight for the top spots." } } }, "zh": { "fullName": "模型排行榜", "factExplain": "按统一评测结果对模型进行排名的榜单。", "humanExplain": "榜单一出，AI 圈立刻进入==放榜模式==：谁冲上前排谁就被围观，分高那位先享受==全场注目礼==。\n\n它常用于选模型和看趋势，但高分不一定等于好用。", "humanExplainDisplay": "榜单一出，\nAI 圈立刻进入==放榜模式==：\n谁冲上前排谁就被围观，\n分高那位先享受==全场注目礼==。\n\n它常用于选模型和看趋势，\n但高分不一定等于好用。", "relationsNarrative": "Benchmark contamination\n如果测试题被模型见过，排行榜成绩就可能好看得不真实。\n\nThird-party-ai-evaluation\n第三方评测能从不同维度补充排行榜，避免只盯单一分数。\n\nLLM-as-a-judge\n有些排行榜会用模型来辅助打分，尤其在主观任务上。\n\nFrontier model\n前沿模型最常在排行榜上被放在一起比较高低。", "relations": { "benchmark-contamination": { "label": "容易被…污染", "note": "题目泄露后，榜单分数会虚高。" }, "third-party-ai-evaluation": { "label": "常参考…结果", "note": "独立测评能补榜单视角不足。" }, "llm-as-a-judge": { "label": "可能借助…打分", "note": "有些榜单会让模型参与评审。" }, "frontier-model": { "label": "常拿…比高低", "note": "前沿模型最常在榜单上厮杀。" } } } } }, { "id": "model-memorization", "name": "Model Memorization", "layer": "L6", "era": "2010s", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "data-privacy" }, { "to": "differential-privacy" }, { "to": "overparameterization" }, { "to": "benchmark-contamination" } ], "track": "today", "i18n": { "en": { "fullName": "Model Memorization", "factExplain": "A mistake where a model repeats training examples instead of making a fresh answer.", "humanExplain": "Model memorization is like a kid who studied by photocopying the answer key. Give it the first line, and it blurts out the whole page, coffee stain included.\n\nThis can leak private data or copyrighted text. It can also make test scores look fake if the model saw the questions before.", "humanExplainDisplay": "Model memorization is like a kid\nwho studied by ==photocopying the answer key==.\nGive it the first line,\nand it blurts out ==the whole page==,\ncoffee stain included.\n\nThis can leak private data\nor copyrighted text.\nIt can also make test scores look fake\nif the model saw the questions before.", "relationsNarrative": "Data-privacy\nModel memorization can reveal private details from training data.\n\nDifferential Privacy\nDifferential Privacy adds noise during training, so memorization is less likely.\n\nOverparameterization\nOverparameterization gives the model more room to memorize examples.\n\nBenchmark contamination\nBenchmark contamination happens when memorized test questions make scores look too high.", "relations": { "data-privacy": { "label": "threatens …", "note": "Memorized training text can come back out with private details." }, "differential-privacy": { "label": "reduced by …", "note": "Differential privacy adds noise during training, so exact copying is harder." }, "overparameterization": { "label": "amplified by …", "note": "Extra model capacity gives more room to memorize samples." }, "benchmark-contamination": { "label": "can cause …", "note": "If the model memorized test questions, scores look too high." } } }, "zh": { "fullName": "模型记忆化", "factExplain": "模型在生成时复现训练样本的现象。", "humanExplain": "模型记忆化像背作文背过头：老师起个头，它连同桌错别字都照念。\n\n它会牵出隐私、版权风险，也会污染评测。", "humanExplainDisplay": "模型记忆化像\n==背作文背过头==：\n老师起个头，\n它连同桌==错别字都照念==。\n\n它会牵出隐私、版权风险，\n也会污染评测。", "relationsNarrative": "Data-privacy\n模型记忆可能让训练数据中的隐私被复现。\n\nDifferential Privacy\n差分隐私通过加噪训练，降低记忆风险。\n\nOverparameterization\n过度参数化会增加模型死记样本的空间。\n\nBenchmark Contamination\n记住评测题会让模型分数看起来虚高。", "relations": { "data-privacy": { "label": "威胁…", "note": "记住训练样本，可能吐出隐私。" }, "differential-privacy": { "label": "用…抑制", "note": "给训练加噪，降低逐字记住。" }, "overparameterization": { "label": "受…放大", "note": "参数太富裕，更容易背答案。" }, "benchmark-contamination": { "label": "导致…", "note": "记住题库会让分数虚高。" } } } } }, { "id": "model-merge", "name": "Model merge", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "open-weights" }, { "to": "fine-tuning" }, { "to": "lora" }, { "to": "base-model" } ], "track": "today", "i18n": { "en": { "fullName": "Model merge", "factExplain": "A way to combine several model weights into one model.", "humanExplain": "Model merge is like pouring three milkshakes into one cup. Chocolate, strawberry, and coffee may work, or taste like a dare.\n\nIt blends models with different styles or skills. You see it often in open-source model tinkering.", "humanExplainDisplay": "Model merge is like pouring\n==three milkshakes== into one cup.\nChocolate, strawberry, and coffee may work,\nor ==taste like a dare==.\n\nIt blends models with different styles or skills.\nYou see it often in open-source model tinkering.", "relationsNarrative": "Open weights\nYou need model weights before you can merge models directly.\n\nFine-tuning\nModels are often fine-tuned first, then merged to combine skills.\n\nLoRA\nMany low-cost workflows merge LoRA adapter effects first.\n\nBase model\nMost merges work best near the same base model.", "relations": { "open-weights": { "label": "needs … to work", "note": "You need the weights before you can merge models directly." }, "fine-tuning": { "label": "often pairs with …", "note": "Teams often fine-tune models first, then merge their strengths." }, "lora": { "label": "can also merge …", "note": "Many workflows merge LoRA adapters first, then make the change permanent." }, "base-model": { "label": "usually stays near …", "note": "Merges work best when the models share a similar base model." } } }, "zh": { "fullName": "模型合并", "factExplain": "把多个模型权重合成为一个模型的方法。", "humanExplain": "这活儿像武侠里拼内功：想把轻功、掌法、护体真气全灌进一人身上，火候不对就容易走火入魔。\n\n常用于开源模型改造，把不同风格或专长揉进同一个模型。", "humanExplainDisplay": "这活儿像武侠里拼==内功==：\n想把轻功、\n掌法、\n护体真气全灌进一人身上，火候不对就容易==走火入魔==。\n\n常用于开源模型改造，\n把不同风格或专长\n揉进同一个模型。", "relationsNarrative": "Open-weights\n能拿到模型权重，才有直接合并的操作空间。\n\nFine-tuning\n模型常先分别微调，再通过合并汇总能力。\n\nLoRA\n很多低成本方案会先合并 LoRA 适配器效果。\n\nBase model\n多数合并更适合同源底座，差太远容易翻车。", "relations": { "open-weights": { "label": "依赖…可操作", "note": "能拿到权重，才谈得上直接合并。" }, "fine-tuning": { "label": "常与…配合", "note": "先分别微调，再合并是常见路线。" }, "lora": { "label": "也可合并…", "note": "很多实践先合并 LoRA，再决定是否烘焙。" }, "base-model": { "label": "通常基于…进行", "note": "多数合并发生在同源底座附近。" } } } } }, { "id": "model-offloading", "name": "Offloading", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-05-31T00:57:33.085Z", "relations": [ { "to": "vram" }, { "to": "local-llm" }, { "to": "quantization" }, { "to": "inference" } ], "track": "today", "i18n": { "en": { "fullName": "Model Offloading", "factExplain": "Temporarily moving some model data to other hardware when memory is too small.", "humanExplain": "Offloading is like cooking a holiday dinner on a tiny kitchen counter. You stash pans in the hallway, then keep cooking.\n\nIt helps run a big model when VRAM is too small. The trade-off is speed. Data has to travel back and forth.", "humanExplainDisplay": "Offloading is like cooking a ==holiday dinner==\non a ==tiny kitchen counter==.\nYou stash pans in the hallway,\nthen keep cooking.\n\nIt helps run a big model\nwhen VRAM is too small.\nThe trade-off is speed.\nData has to travel back and forth.", "relationsNarrative": "VRAM\nOffloading moves some model data elsewhere when VRAM is too small.\n\nLocal-LLM\nOffloading can help a local device run a model it could not fit before.\n\nQuantization\nQuantization cuts size, and Offloading moves data to another place.\n\nInference\nOffloading often happens during inference and trades speed for the chance to run.", "relations": { "vram": { "label": "eases low …", "note": "When VRAM is too small, offloading moves some data elsewhere." }, "local-llm": { "label": "helps … fit", "note": "It gives local devices a better shot at running bigger models." }, "quantization": { "label": "often works with …", "note": "Both reduce hardware pressure, but in different ways." }, "inference": { "label": "happens during …", "note": "During inference, it trades speed for a model that can run." } } }, "zh": { "fullName": "模型卸载", "factExplain": "把部分模型数据临时挪到别的硬件存放。", "humanExplain": "模型卸载像小户型收纳：GPU 放不下，就把不常用的先塞进储物间，现用现搬。\n\n它常用于本地跑大模型，省显存，但会拖慢速度。", "humanExplainDisplay": "模型卸载像==小户型收纳==：\nGPU 放不下，\n就把不常用的先塞进储物间，\n==现用现搬==。\n\n它常用于本地跑大模型，\n省显存，\n但会拖慢速度。", "relationsNarrative": "VRAM\nOffloading 常在显存不够时，把部分模型数据挪到别处。\n\nLocal-LLM\n它能让本地设备有机会运行原本塞不下的大模型。\n\nQuantization\nQuantization 是压缩体积，Offloading 是挪存放位置，常一起用。\n\nInference\nOffloading 多发生在推理阶段，用速度换取可运行性。", "relations": { "vram": { "label": "缓解…不够", "note": "显存装不下时，会把部分数据挪走。" }, "local-llm": { "label": "帮…挤上车", "note": "让本地设备更有机会跑更大的模型。" }, "quantization": { "label": "常与…配合", "note": "两者都在想办法降低硬件压力。" }, "inference": { "label": "发生在…阶段", "note": "模型运行时常靠它换取可用性。" } } } } }, { "id": "model-parallelism", "name": "Model parallelism", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "vram" }, { "to": "model-offloading" }, { "to": "inference" } ], "track": "today", "i18n": { "en": { "fullName": "Model parallelism", "factExplain": "A way to split one model across several devices to work together.", "humanExplain": "Think of a huge sofa stuck at the stairs. You split it into pieces. Then each friend carries one piece upstairs.\n\nModel parallel does this for very big AI models. It helps when one GPU does not have enough VRAM, in training or inference.", "humanExplainDisplay": "Think of a ==huge sofa==\nstuck at the stairs.\nYou ==split it into pieces==.\nThen each friend carries\none piece upstairs.\n\nModel parallel does this for very big AI models.\nIt helps when one GPU does not have enough VRAM,\nin training or inference.", "relationsNarrative": "GPU\nModel parallel splits the model and places the parts on several GPUs.\n\nVRAM\nPeople often use it when one GPU's VRAM cannot fit the whole model.\n\nOffloading\nOffloading is the other fix for a too-big model — it moves parts away instead of splitting.\n\nInference\nVery large models often need it to run inference smoothly.", "relations": { "gpu": { "label": "splits the model across …", "note": "The model is spread across several GPUs that compute together." }, "vram": { "label": "helps when … is too small", "note": "People use it when one GPU's VRAM cannot hold the whole model." }, "model-offloading": { "label": "replaces or works with …", "note": "Both fix the same size problem. One splits across GPUs. One moves parts elsewhere." }, "inference": { "label": "used for large-model …", "note": "It also helps very large models run during inference." } } }, "zh": { "fullName": "Model parallelism｜模型并行", "factExplain": "把一个模型拆到多块设备上协同计算的方法。", "humanExplain": "模型并行像搬家请兄弟分扛大衣柜：一人扛不动，就拆开大家抬。\n\n它让超大模型能在多 GPU 上训练和推理，但通信会拖后腿。", "humanExplainDisplay": "模型并行像搬家请兄弟==分扛大衣柜==：\n一人扛不动，\n就==拆开大家抬==。\n\n它让超大模型能在多GPU上训练和推理，\n但通信会拖后腿。", "relationsNarrative": "GPU\nModel parallelism 会把模型切分后放到多张 GPU 上一起算。\n\nVRAM\n当单卡 VRAM 放不下整个模型时，常用它来分摊容量压力。\n\nModel-offloading\n两者都在解决装不下的问题，只是一个拆多卡，一个往别处挪。\n\nInference\n超大模型做 Inference 时，也常靠它才能顺利运行。", "relations": { "gpu": { "label": "把模型拆给…", "note": "模型会分布到多张 GPU 协同计算。" }, "vram": { "label": "缓解…不够", "note": "当单卡显存装不下模型时常用它。" }, "model-offloading": { "label": "替代或配合…", "note": "装不下时，可拆多卡或部分挪到别处。" }, "inference": { "label": "用于大模型…", "note": "推理阶段也常靠它运行超大模型。" } } } } }, { "id": "model-regression", "name": "Model regression", "layer": "L5", "sublayer": "architecture", "era": "2010s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "llmops" }, { "to": "third-party-ai-evaluation" }, { "to": "ai-qa-testing" }, { "to": "fine-tuning" } ], "track": "today", "i18n": { "en": { "fullName": "Model regression", "factExplain": "A model update fixes something, but an old skill gets worse.", "humanExplain": "Model regression is like a game update adding laser swords. Then the jump button breaks.\n\nYou meet it after model updates or fine-tuning. Teams rerun old tests before launch to catch it.", "humanExplainDisplay": "Model regression is like a game update\nadding ==laser swords==.\nThen the ==jump button breaks==.\n\nYou meet it after model updates\nor fine-tuning.\nTeams rerun old tests before launch\nto catch it.", "relationsNarrative": "LLMOps\nLLMOps uses version monitoring to spot model regression.\n\nThird-party AI evaluation\nThird-party AI evaluation can catch regressions the team missed.\n\nAI QA Testing\nAI QA Testing reruns old cases before launch to block regressions.\n\nFine-tuning\nFine-tuning can improve a new skill and hurt an old one.", "relations": { "llmops": { "label": "is monitored by …", "note": "Version monitoring can spot old skills getting worse." }, "third-party-ai-evaluation": { "label": "is checked by …", "note": "Outside tests can find regressions your own tests missed." }, "ai-qa-testing": { "label": "is blocked with …", "note": "Pre-launch tests rerun old cases and catch broken skills." }, "fine-tuning": { "label": "can come from …", "note": "Fine-tuning may improve one skill and hurt another." } } }, "zh": { "fullName": "模型退化（回归缺陷）", "factExplain": "模型更新后在旧能力上变差的现象。", "humanExplain": "模型回归像补作业：新题刚会做，期中错题又忘了怎么解。\n\n它要求上线前跑旧题，防止新模型把老本事改坏。", "humanExplainDisplay": "模型回归像补作业：\n==新题刚会做==，\n==期中错题==又忘了怎么解。\n\n它要求上线前跑旧题，\n防止新模型，\n把老本事改坏。", "relationsNarrative": "LLMOps\nLLMOps 通过版本监控发现模型能力倒退。\n\nThird-party AI evaluation\n第三方评测能揭出团队自测漏掉的退步。\n\nAI QA Testing\nAI QA Testing 用旧用例拦截上线前退化。\n\nFine-tuning\nFine-tuning 可能提升新能力，也弄坏旧能力。", "relations": { "llmops": { "label": "由…监控", "note": "版本监控能及时发现能力倒退。" }, "third-party-ai-evaluation": { "label": "靠…复查", "note": "外部评测能揭出自测漏掉的退步。" }, "ai-qa-testing": { "label": "用…拦截", "note": "上线前测试旧用例，防止翻车。" }, "fine-tuning": { "label": "可能来自…", "note": "微调补一块，也可能弄坏另一块。" } } } } }, { "id": "model-routing", "name": "Routing", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "inference" }, { "to": "agent" }, { "to": "llmops" } ], "track": "today", "seo": { "en": { "title": "What Is AI Model Routing? Triage for LLM Requests", "description": "Model routing sends each request to the model that fits — easy stuff to cheap models, hard cases to the experts, like hospital triage. A plain-English explainer." }, "zh": { "title": "模型路由是什么?给请求做分诊的省钱策略,一文看懂 — AI Rookies", "description": "小感冒别挂院士号:按成本、速度、难度把请求分给合适的模型。多模型服务的常见做法,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Model Routing", "factExplain": "A strategy that sends each request to the model that fits it best.", "humanExplain": "Do not call the top chef for every snack. Model routing is like a host at the door. Easy orders go fast. Hard ones go to the expert.\n\nIt picks the path inside a multi-model product. It helps balance quality, cost, and speed.", "humanExplainDisplay": "Do not call the ==top chef==\nfor every snack.\nModel routing is like a ==host at the door==.\nEasy orders go fast.\nHard ones go to the expert.\n\nIt picks the path inside a multi-model product.\nIt helps balance quality, cost, and speed.", "relationsNarrative": "LLM\nModel routing sends each request to the LLM that fits it best.\n\nInference\nIt decides which inference path a request will take.\n\nAgent\nAn Agent often switches models across different steps.\n\nLLMOps\nModel routing is part of multi-model deployment and cost control.", "relations": { "llm": { "label": "routes requests to …", "note": "It sends different tasks to the LLM that fits best." }, "inference": { "label": "happens around …", "note": "Routing decides which inference path a request will take." }, "agent": { "label": "picks models for …", "note": "An Agent may switch models for different steps." }, "llmops": { "label": "is part of … optimization", "note": "It is a common way to run many models better." } } }, "zh": { "fullName": "Model routing｜模型路由", "factExplain": "按任务把请求分配给不同模型的策略。", "humanExplain": "它像医院分诊：小感冒别挂院士号，疑难杂症再请专家上场。\n\n它常用于多模型服务，按成本、速度和难度分配请求。", "humanExplainDisplay": "它像==医院分诊==：\n==小感冒别挂院士号==，\n疑难杂症再请专家上场。\n\n它常用于多模型服务，\n按成本、速度和难度分配请求。", "relationsNarrative": "LLM\nModel routing 会把请求分配给更合适的 LLM。\n\nInference\n它决定一次请求该进入哪条推理路径。\n\nAgent\nAgent 执行不同环节时，常会切换不同模型。\n\nLlmops\nModel routing 是多模型部署与成本优化的一部分。", "relations": { "llm": { "label": "为…分流请求", "note": "把不同任务分给更合适的 LLM。" }, "inference": { "label": "发生在…前后", "note": "路由决定请求交给哪条推理路径。" }, "agent": { "label": "给…挑模型", "note": "Agent 可按步骤切换不同模型执行。" }, "llmops": { "label": "属于…优化", "note": "它是多模型上线运营的常见手段。" } } } } }, { "id": "model-scaffolding", "name": "Model scaffolding", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "agent-harness" }, { "to": "agent" }, { "to": "function-call" }, { "to": "structured-output" } ], "track": "today", "i18n": { "en": { "fullName": "Model scaffolding", "factExplain": "A design that wraps an AI model with steps, tools, and rules.", "humanExplain": "Model scaffolding is the clipboard for a school field trip. The AI still walks, but the teacher whistle keeps it on track.\n\nIt breaks a big task into steady steps. You meet it in Agents. You also meet it in coding helpers and support bots.", "humanExplainDisplay": "Model scaffolding is the ==clipboard for a school field trip==.\nThe AI still walks,\nbut the ==teacher whistle== keeps it on track.\n\nIt breaks a big task into steady steps.\nYou meet it in Agents.\nYou also meet it in coding helpers\nand support bots.", "relationsNarrative": "Agent harness\nAn Agent harness schedules the scaffold’s steps.\n\nAgent\nAn Agent uses scaffolding to split tasks and act step by step.\n\nFunction-calling\nFunction-calling lets scaffolded steps do real work.\n\nStructured output\nStructured output makes each step easier to check and pass on.", "relations": { "agent-harness": { "label": "becomes runnable in …", "note": "An Agent harness turns the scaffold into a working shell." }, "agent": { "label": "supports … actions", "note": "Scaffolding helps an Agent plan and act step by step." }, "function-call": { "label": "connects tools through …", "note": "Function-calling turns planned steps into real actions." }, "structured-output": { "label": "controls format with …", "note": "Structured output makes each step easy to check." } } }, "zh": { "fullName": "模型脚手架", "factExplain": "用外部流程、工具和约束包住模型的系统设计。", "humanExplain": "模型脚手架像煎饼摊的摊位动线：面糊、鸡蛋、薄脆排好，师傅不慌也少漏。\n\n它把大任务拆成稳步骤，常用于智能体、代码助手和客服。", "humanExplainDisplay": "模型脚手架像煎饼摊的\n==摊位动线==：\n面糊、鸡蛋、薄脆排好，\n师傅==不慌也少漏==。\n\n它把大任务拆成稳步骤，\n常用于智能体、\n代码助手和客服。", "relationsNarrative": "Agent harness\n它是脚手架的工程外壳，负责调度步骤。\n\nAgent\n智能体常靠脚手架拆任务、记状态、调用工具。\n\nFunction-calling\n工具调用让脚手架安排的步骤能真正执行。\n\nStructured output\n结构化输出让每一步结果更容易校验和传递。", "relations": { "agent-harness": { "label": "落进…里", "note": "它把脚手架变成可运行外壳。" }, "agent": { "label": "支撑…行动", "note": "它让智能体能分步规划和执行。" }, "function-call": { "label": "接上…工具", "note": "工具调用把步骤变成真实动作。" }, "structured-output": { "label": "约束…格式", "note": "结构化输出让每步结果可检查。" } } } } }, { "id": "model-uncertainty", "name": "Model uncertainty", "layer": "L4", "era": "2023", "publishedAt": "2026-05-28T15:58:23.418Z", "relations": [ { "to": "hallucination" }, { "to": "temperature" }, { "to": "alignment" }, { "to": "rag" } ], "track": "today", "i18n": { "en": { "fullName": "Model uncertainty", "factExplain": "How sure a model seems about its own answer.", "humanExplain": "Model uncertainty is the AI's quiz-show wobble. It slams the buzzer, then whispers, “maybe penguins?”\n\nIt tells you how much to trust an answer. It matters in Q&A and medical tools. It matters in automated decisions too.", "humanExplainDisplay": "Model uncertainty is the AI's ==quiz-show wobble==.\nIt ==slams the buzzer==,\nthen whispers, “maybe penguins?”\n\nIt tells you how much to trust an answer.\nIt matters in Q&A and medical tools.\nIt matters in automated decisions too.", "relationsNarrative": "Hallucination\nHigher model uncertainty means more risk of confident nonsense.\n\nTemperature\nTemperature can make answers more random, so the model may seem less sure.\n\nAlignment\nAlignment helps the model hold back when it is not sure.\n\nRAG\nRAG gives the model outside facts, so uncertainty often goes down.", "relations": { "hallucination": { "label": "helps spot … risk", "note": "High uncertainty means you should watch for confident nonsense." }, "temperature": { "label": "is affected by …", "note": "Higher Temperature often makes answers less steady." }, "alignment": { "label": "is a signal for …", "note": "Alignment teaches the model to hold back when unsure." }, "rag": { "label": "can be lowered by …", "note": "RAG gives sources, so answers often get steadier." } } }, "zh": { "fullName": "模型不确定性", "factExplain": "模型对自己输出把握程度的表现。", "humanExplain": "看它回话那股劲儿就知道：有时像学霸秒答，有时像硬着头皮编到后半句开始发虚。\n\n它决定答案有多能信，常影响问答、医疗和自动化决策。", "humanExplainDisplay": "看它回话那股劲儿就知道：\n有时像==学霸秒答==，\n有时像硬着头皮\n编到后半句\n开始==发虚==。\n\n它决定答案有多能信，\n常影响问答、\n医疗和自动化决策。", "relationsNarrative": "Hallucination\n模型不确定性越高，往往越需要警惕它在一本正经地胡说。\n\nTemperature\nTemperature 会影响输出的发散程度，也会改变模型看起来有多笃定。\n\nAlignment\nAlignment 不只关心答什么，也关心模型该在没把握时适当收住。\n\nRAG\nRAG 给模型补充外部资料后，很多问题上的不确定性会下降。", "relations": { "hallucination": { "label": "帮助识别…风险", "note": "不确定性高时，更该警惕胡编。" }, "temperature": { "label": "会受…影响", "note": "温度越高，输出通常越飘忽。" }, "alignment": { "label": "是…关注信号", "note": "让模型知道何时该保留意见。" }, "rag": { "label": "可被…降低", "note": "有资料可查时，答复通常更稳。" } } } } }, { "id": "momentum", "name": "Momentum", "layer": "L2", "era": "1964", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "sgd" }, { "to": "gradient-descent" }, { "to": "adam" } ], "track": "history", "i18n": { "en": { "fullName": "Momentum Optimization", "factExplain": "An optimizer trick that uses past gradients to speed up model updates.", "humanExplain": "Momentum is like pushing a grocery cart. Once it rolls straight, tiny bumps in the cereal aisle do not boss it around.\n\nIn neural network training, it helps updates move faster and wobble less. You often see it with SGD, and inside Adam.", "humanExplainDisplay": "Momentum is like pushing a ==grocery cart==.\nOnce it rolls straight,\ntiny bumps in the ==cereal aisle== do not boss it around.\n\nIn neural network training,\nit helps updates move faster and wobble less.\nYou often see it with SGD,\nand inside Adam.", "relationsNarrative": "SGD\nMomentum often upgrades SGD, so it updates faster and more smoothly.\n\nGradient Descent\nMomentum improves Gradient Descent by reducing zigzag movement.\n\nAdam\nAdam builds on momentum and also adjusts step sizes.", "relations": { "sgd": { "label": "speeds up …", "note": "Momentum adds past update direction to SGD." }, "gradient-descent": { "label": "smooths … updates", "note": "It helps Gradient Descent move forward with less zigzagging." }, "adam": { "label": "came before …", "note": "Adam uses the core idea of momentum." } } }, "zh": { "fullName": "动量优化", "factExplain": "一种利用历史梯度加速参数更新的优化方法。", "humanExplain": "动量像下棋起势：这步既然方向对，后面就顺着攻，别被眼前一两手小波动带偏。\n\n它常用来加快收敛、减少震荡，尤其适合深度网络训练。", "humanExplainDisplay": "动量像下棋起势：\n这步既然==方向对==，\n后面就顺着攻，\n别被==一两手小波动==带偏。\n\n它常用来加快收敛、\n减少震荡，\n尤其适合深度网络训练。", "relationsNarrative": "SGD\n动量常作为 SGD 的增强版，帮它更新得更稳更快。\n\nGradient Descent\n它改进梯度下降的更新轨迹，减少锯齿式震荡。\n\nAdam\nAdam 吸收了动量思想，并叠加自适应学习率。", "relations": { "sgd": { "label": "作为…的加速版", "note": "常在 SGD 上加入历史更新方向。" }, "gradient-descent": { "label": "改进…更新", "note": "它让参数更新少些摇摆、更快前进。" }, "adam": { "label": "是…的重要前身", "note": "Adam 继承了动量这一核心思路。" } } } } }, { "id": "monte-carlo-tree-search", "name": "MCTS", "layer": "L2", "era": "2006", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "graph-search" }, { "to": "heuristic-search" }, { "to": "reinforcement-learning" }, { "to": "markov-decision-process" } ], "track": "history", "i18n": { "en": { "fullName": "Monte Carlo Tree Search", "factExplain": "A search method for choosing actions through many pretend tries.", "humanExplain": "MCTS is like a chess player with a tiny movie studio. It films many pretend games, then picks the move with the best endings.\n\nIt helps AI plan many steps ahead. You see it in game AI and tough choice problems.", "humanExplainDisplay": "MCTS is like a chess player\nwith a ==tiny movie studio==.\nIt films many ==pretend games==,\nthen picks the move with the best endings.\n\nIt helps AI plan many steps ahead.\nYou see it in game AI\nand tough choice problems.", "relationsNarrative": "Graph Search\nMCTS is Graph Search on a tree of possible moves.\n\nHeuristic Search\nHeuristic Search uses quick guesses, but MCTS uses many pretend tries.\n\nRL\nMCTS can give RL stronger action choices and better planning.\n\nMDP\nMCTS can help an MDP find better action paths.", "relations": { "graph-search": { "label": "is a kind of …", "note": "It searches a tree by trying branches again and again." }, "heuristic-search": { "label": "works alongside …", "note": "Heuristic Search guesses scores. MCTS leans on pretend tries." }, "reinforcement-learning": { "label": "often helps …", "note": "MCTS can help RL choose stronger actions." }, "markov-decision-process": { "label": "can solve choices in …", "note": "MCTS can search for better action paths in an MDP." } } }, "zh": { "fullName": "Monte Carlo Tree Search｜蒙特卡洛树搜索", "factExplain": "一种靠反复模拟来选择行动的搜索方法。", "humanExplain": "像下棋时先在脑内偷演几步：这手走下去是送子还是将军，试得多了就更敢落子。\n\n常用于多步规划，在分支很多时靠模拟逐步选出更优路径。", "humanExplainDisplay": "像下棋时先在脑内==偷演几步==：\n这手走下去是==送子还是将军==，\n试得多了就更敢落子。\n\n常用于多步规划，\n在分支很多时靠模拟\n逐步选出更优路径。", "relationsNarrative": "Graph Search\n它是在树结构上展开分支并反复评估的搜索方法。\n\nHeuristic Search\n启发式搜索靠规则估价，它更依赖模拟结果选路。\n\nReinforcement Learning\n它常为强化学习提供更强的动作选择与规划能力。\n\nMarkov Decision Process\n它可用于序列决策中近似寻找更优行动路径。", "relations": { "graph-search": { "label": "属于…一类", "note": "它是在树结构上反复试探的搜索法。" }, "heuristic-search": { "label": "与…互补", "note": "一个靠估价，一个更靠模拟试走。" }, "reinforcement-learning": { "label": "常和…配合", "note": "可用模拟搜索帮强化学习选动作。" }, "markov-decision-process": { "label": "可用于求解决策", "note": "适合序列决策里的行动选择问题。" } } } } }, { "id": "moravec-s-paradox", "name": "Moravec's Paradox", "layer": "L1", "era": "1988", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "embodied-ai" }, { "to": "computer-vision" }, { "to": "agi" } ], "track": "history", "i18n": { "en": { "fullName": "Moravec's Paradox", "factExplain": "High-level reasoning can be easy for AI, while seeing and moving are hard.", "humanExplain": "Moravec's Paradox is AI getting an A in calculus. Then it loses a fight with a doorknob.\n\nIt explains why robots are hard. It also explains why seeing and moving in the real world are hard.", "humanExplainDisplay": "Moravec's Paradox is AI getting\nan ==A in calculus==.\nThen it ==loses a fight with a doorknob==.\n\nIt explains why robots are hard.\nIt also explains why seeing and moving\nin the real world are hard.", "relationsNarrative": "Robotics\nMoravec's Paradox explains why robots are hard to make reliable in real places.\n\nEmbodied AI\nMoravec's Paradox shows intelligence needs a body, senses, and action.\n\nComputer Vision\nMoravec's Paradox points out real-world seeing is harder than it looks.\n\nAGI\nMoravec's Paradox challenges the idea that reasoning alone gets us close to AGI.", "relations": { "robotics": { "label": "explains why … is hard", "note": "Robots struggle with tiny details in seeing and moving." }, "embodied-ai": { "label": "shows … needs a body", "note": "Body-based tasks reveal hard problems in common sense and control." }, "computer-vision": { "label": "points out … is hard", "note": "Seeing the real world means more than naming objects." }, "agi": { "label": "challenges … hopes", "note": "General intelligence cannot be just paper reasoning." } } }, "zh": { "fullName": "莫拉维克悖论", "factExplain": "指出高层推理易自动化，感知运动反而难。", "humanExplain": "莫拉维克悖论：AI 会考清华题，去菜场挑葱却躲不开电驴。\n\n解释机器人、视觉和具身任务，为何难落地。", "humanExplainDisplay": "莫拉维克悖论：\nAI 会==考清华题==，\n去菜场挑葱，\n却躲不开==电驴==。\n\n解释机器人、视觉和具身任务，\n为何难落地。", "relationsNarrative": "Robotics\n它解释了为什么机器人在现实环境中难做稳。\n\nEmbodied AI\n它提醒智能离不开身体、感知和动作。\n\nComputer Vision\n它点出看懂真实世界比人想的更难。\n\nAGI\n它挑战“会推理就接近通用智能”的想象。", "relations": { "robotics": { "label": "解释…为何难", "note": "机器人难在感知和动作细节。" }, "embodied-ai": { "label": "提醒…需要身体", "note": "具身任务暴露常识与控制难题。" }, "computer-vision": { "label": "点出…的难度", "note": "看懂真实世界不只是识别物体。" }, "agi": { "label": "挑战…想象", "note": "通用智能不能只会纸面推理。" } } } } }, { "id": "mujoco", "name": "MuJoCo", "layer": "L5", "sublayer": "product", "era": "2012", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "openai-gym" }, { "to": "reinforcement-learning" }, { "to": "robotics" }, { "to": "model-based-reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Multi-Joint dynamics with Contact", "factExplain": "A physics simulator for testing robot movement and control.", "humanExplain": "MuJoCo is a padded gym for robots. They can trip, face-plant, and break only pretend knees.\n\nIt trains robot control and RL in simulation. You burn computer time first, not pricey robot parts.", "humanExplainDisplay": "MuJoCo is a ==padded gym== for robots.\nThey can trip, face-plant,\nand break only ==pretend knees==.\n\nIt trains robot control and RL in simulation.\nYou burn computer time first,\nnot pricey robot parts.", "relationsNarrative": "OpenAI Gym\nOpenAI Gym used MuJoCo for classic continuous control environments.\n\nRL\nRL agents can use MuJoCo to learn by cheap trial and error.\n\nRobotics\nMuJoCo simulates robot joints, contact, and motion.\n\nMBRL\nMBRL often uses MuJoCo to test physics predictions.", "relations": { "openai-gym": { "label": "provides environments for …", "note": "OpenAI Gym used MuJoCo for classic continuous control tasks." }, "reinforcement-learning": { "label": "gives … a practice field", "note": "RL agents can fail again and again in simulation." }, "robotics": { "label": "simulates … motion", "note": "Robots can practice joints and contact in a virtual world." }, "model-based-reinforcement-learning": { "label": "supports … modeling", "note": "Accurate physics helps predict the next move." } } }, "zh": { "fullName": "Multi-Joint dynamics with Contact（多关节接触动力学）", "factExplain": "一个用于机器人控制的物理仿真引擎。", "humanExplain": "MuJoCo是机器人练武的木人桩：招式随便摔，断的是仿真骨头，不是真零件。\n\n用于训练机器人控制和强化学习，先烧算力，少烧硬件。", "humanExplainDisplay": "MuJoCo是机器人练武的\n==木人桩==：\n招式随便摔，\n断的是==仿真骨头==，不是真零件。\n\n用于训练机器人控制和强化学习，\n先烧算力，\n少烧硬件。", "relationsNarrative": "OpenAI Gym\nOpenAI Gym 曾用它提供经典连续控制环境。\n\nReinforcement Learning\n强化学习智能体可在其中低成本反复试错。\n\nRobotics\n它能模拟机器人关节、接触和运动。\n\nModel-based Reinforcement Learning\n模型强化学习常借它验证动力学预测。", "relations": { "openai-gym": { "label": "提供…环境", "note": "Gym 曾把它做成经典控制任务。" }, "reinforcement-learning": { "label": "为…提供训练场", "note": "智能体可在仿真中反复试错。" }, "robotics": { "label": "模拟…运动", "note": "机器人先在虚拟世界练动作。" }, "model-based-reinforcement-learning": { "label": "支撑…建模", "note": "精确动力学便于预测下一步。" } } } } }, { "id": "multi-agent-system", "name": "MAS", "layer": "L5", "sublayer": "architecture", "era": "2024", "publishedAt": "2026-05-31T00:57:33.081Z", "relations": [ { "to": "agent" }, { "to": "agent-harness" }, { "to": "function-call" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "Multi-agent system", "factExplain": "A system where several Agents split a task and work together.", "humanExplain": "MAS is an AI group project. One kid researches. One makes slides. One keeps asking if this is a bad idea.\n\nIt helps with complex workflows, like an automated team. But more roles can also mean more chaos.", "humanExplainDisplay": "MAS is an ==AI group project==.\nOne kid researches.\nOne makes slides.\nOne keeps asking\nif this is a ==bad idea==.\n\nIt helps with complex workflows,\nlike an automated team.\nBut more roles can also mean\nmore chaos.", "relationsNarrative": "Agent\nA multi-agent system is several Agents splitting the work.\n\nAgent harness\nAn Agent harness manages the workflow between Agents.\n\nFunction-calling\nMulti-agent systems often use Function-call to do specific actions.\n\nHuman-in-the-loop\nHuman-in-the-loop adds review at key steps.", "relations": { "agent": { "label": "is made of …", "note": "A multi-agent system is several Agents working together." }, "agent-harness": { "label": "is coordinated by …", "note": "An Agent harness manages the roles and workflow." }, "function-call": { "label": "acts through …", "note": "Each Agent often uses Function-call to take real actions." }, "human-in-the-loop": { "label": "uses … as backup", "note": "Human review helps stop the team from going off track." } } }, "zh": { "fullName": "Multi-agent system｜多智能体系统", "factExplain": "由多个 Agent 分工协作完成任务的系统。", "humanExplain": "一个 AI 忙不过来？那就拉个群：有人查资料，有人写方案，还有人专门负责唱反调。\n\n适合复杂协作流程，常见于自动化团队；但角色越多，越容易跑偏。", "humanExplainDisplay": "一个 AI 忙不过来？\n那就拉个==群==：\n有人查资料，\n有人写方案，\n还有人专门负责\n==唱反调==。\n\n适合复杂协作流程，\n常见于自动化团队；\n但角色越多，\n越容易跑偏。", "relationsNarrative": "Agent\n多智能体系统本质上是多个 Agent 分工协作。\n\nAgent harness\nAgent harness 常用来管理多 Agent 的流程与协作。\n\nFunction-calling\n多 Agent 往往靠 Function-call 执行具体动作。\n\nHuman-in-the-loop\n关键环节加入人类审核，能给多 Agent 协作兜底。", "relations": { "agent": { "label": "由…组成", "note": "多智能体系统本质上是多个 Agent 协作。" }, "agent-harness": { "label": "靠…编排", "note": "Agent harness 常用来管理角色与流程。" }, "function-call": { "label": "用…执行动作", "note": "各个 Agent 往往靠工具调用落地任务。" }, "human-in-the-loop": { "label": "配…兜底", "note": "关键节点常需人类审核防止集体跑偏。" } } } } }, { "id": "multi-armed-bandit", "name": "Bandit", "layer": "L2", "era": "1952", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "exploration-exploitation" }, { "to": "thompson-sampling" }, { "to": "upper-confidence-bound" }, { "to": "reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Multi-Armed Bandit Problem", "factExplain": "A problem about when to try new choices and when to trust known winners.", "humanExplain": "A Bandit problem is like a row of mystery gumball machines. One machine is the candy king, but your quarters are limited.\n\nIt helps systems test choices while still earning rewards. You meet it in recommendations, ads, and online tests.", "humanExplainDisplay": "A Bandit problem is like\na row of ==mystery gumball machines==.\nOne machine is the ==candy king==,\nbut your quarters are limited.\n\nIt helps systems test choices\nwhile still earning rewards.\nYou meet it in recommendations,\nads, and online tests.", "relationsNarrative": "Exploration-Exploitation Tradeoff\nBandit is the classic practice problem for trying new paths versus using safe ones.\n\nThompson Sampling\nThompson Sampling is one common way to solve a Bandit problem.\n\nUCB\nUCB uses a hopeful upper score to decide which arm deserves more tries.\n\nRL\nBandit is often seen as the simplest starter scene in RL.", "relations": { "exploration-exploitation": { "label": "models …", "note": "Bandit is the classic problem of trying new options versus using known winners." }, "thompson-sampling": { "label": "can be solved with …", "note": "Thompson Sampling is a common chance-based way to solve Bandit problems." }, "upper-confidence-bound": { "label": "can be solved with …", "note": "UCB tries the option with the best hopeful score." }, "reinforcement-learning": { "label": "is a starter form of …", "note": "Bandit is often seen as the simple first step into RL." } } }, "zh": { "fullName": "多臂老虎机问题", "factExplain": "研究如何在探索新选项与利用已知收益间做决策的问题。", "humanExplain": "追剧时，老剧最稳，新剧可能封神；今晚只够看一集，你得边试边押宝。\n\n它用于推荐、广告和实验分流，动态把流量给更优选项。", "humanExplainDisplay": "追剧时，==老剧最稳==，\n新剧可能==封神==；\n今晚只够看一集，\n你得边试边押宝。\n\n它用于推荐、广告\n和实验分流，\n动态把流量给更优选项。", "relationsNarrative": "Exploration-Exploitation Tradeoff\n它就是“试新路”和“走熟路”的经典练习题。\n\nThompson Sampling\nThompson Sampling 是求解它的代表性方法之一。\n\nUpper-Confidence-Bound\nUCB 用乐观上界来决定该多试哪一臂。\n\nReinforcement-learning\n它常被视为强化学习里最基础的入门场景。", "relations": { "exploration-exploitation": { "label": "刻画…权衡", "note": "它就是这类取舍的经典问题。" }, "thompson-sampling": { "label": "可用…求解", "note": "它是常见的概率式策略之一。" }, "upper-confidence-bound": { "label": "可用…求解", "note": "它用乐观估计来平衡尝试与收益。" }, "reinforcement-learning": { "label": "属于…前菜", "note": "它常被看作强化学习的简化版。" } } } } }, { "id": "multi-gpu-inference", "name": "Multi-GPU inference", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "model-parallelism" }, { "to": "llm-inference-engine" }, { "to": "vram" } ], "track": "today", "i18n": { "en": { "fullName": "Multi-GPU inference", "factExplain": "Using several GPUs together to run a model and answer requests.", "humanExplain": "Multi-GPU inference is like moving a giant couch with one friend. Add more friends, and nobody becomes pancake.\n\nCloud apps and company AI use it. It helps huge models fit and reply faster.", "humanExplainDisplay": "Multi-GPU inference is like\n==moving a giant couch== with one friend.\nAdd more friends,\nand ==nobody becomes pancake==.\n\nCloud apps and company AI use it.\nIt helps huge models fit\nand reply faster.", "relationsNarrative": "GPU\nMulti-GPU inference splits one inference job across several GPUs.\n\nModel parallelism\nModel parallelism is a common way to split the model across GPUs.\n\nInference engine\nThe inference engine schedules the GPU work and communication.\n\nVRAM\nMulti-GPU inference can combine VRAM so a larger model fits.", "relations": { "gpu": { "label": "splits work across …", "note": "Several GPUs share the work for one inference job." }, "model-parallelism": { "label": "often uses …", "note": "Model parallelism can split a huge model across GPUs." }, "llm-inference-engine": { "label": "is scheduled by …", "note": "The inference engine assigns compute and messages between GPUs." }, "vram": { "label": "combines … capacity", "note": "Multiple GPUs can provide enough VRAM for a larger model." } } }, "zh": { "fullName": "多 GPU 推理", "factExplain": "使用多块 GPU 协同执行模型推理。", "humanExplain": "多卡推理像龙舟队划船：一人划不动大船，多人同频才跑得快。\n\n用于云端和企业部署，让大模型装得下、答得快。", "humanExplainDisplay": "多卡推理像\n==龙舟队划船==：\n一人划不动大船，\n==多人同频==才跑得快。\n\n用于云端和企业部署，\n让大模型装得下、答得快。", "relationsNarrative": "GPU\n多卡推理把一次推理分给多块 GPU 协同完成。\n\nModel Parallelism\n模型并行是多卡推理常见的拆分方式。\n\nInference Engine\n推理引擎负责调度多卡计算和通信。\n\nVRAM\n多卡常用来凑显存，装下单卡放不下的模型。", "relations": { "gpu": { "label": "把任务分给…", "note": "多块 GPU 协同扛住推理。" }, "model-parallelism": { "label": "常用…拆模型", "note": "模型太大时，可切到多卡上跑。" }, "llm-inference-engine": { "label": "由…调度", "note": "推理引擎负责分配计算与通信。" }, "vram": { "label": "合并…容量", "note": "多卡常为装下更大的模型。" } } } } }, { "id": "multi-token-prediction", "name": "MTP", "layer": "L2", "era": "2024", "publishedAt": "2026-05-30T03:10:23.229Z", "relations": [ { "to": "token" }, { "to": "inference" }, { "to": "llm-inference-engine" }, { "to": "continuous-batching" } ], "track": "today", "i18n": { "en": { "fullName": "Multi-Token Prediction", "factExplain": "A way for a model to predict several tokens in one step.", "humanExplain": "Old-school AI drops words like a slow gum machine. MTP turns the knob and gets a small handful.\n\nIt helps AI answer faster during inference. You meet it inside large-model speed tricks.", "humanExplainDisplay": "Old-school AI drops words\nlike a ==slow gum machine==.\nMTP turns the knob\nand gets ==a small handful==.\n\nIt helps AI answer faster\nduring inference.\nYou meet it inside\nlarge-model speed tricks.", "relationsNarrative": "Token\nMTP predicts several tokens at once instead of one at a time.\n\nInference\nMTP is mainly used during inference to make generation faster.\n\nInference engine\nMTP often needs an inference engine to run well in real systems.\n\nContinuous batching\nContinuous batching handles the waiting line, and MTP makes each step bigger.", "relations": { "token": { "label": "predicts several … at once", "note": "MTP changes one-by-one token generation into group generation." }, "inference": { "label": "speeds up …", "note": "MTP is a speed trick used during inference." }, "llm-inference-engine": { "label": "is run by …", "note": "An inference engine usually schedules MTP in real use." }, "continuous-batching": { "label": "can speed up with …", "note": "Both help online AI generation run faster." } } }, "zh": { "fullName": "多 token 预测", "factExplain": "让模型一次预测多个 token 的生成方式。", "humanExplain": "以前是一颗一颗往外蹦字，现在干脆==一把吐几颗==，像输入法提前读懂你，直接==整段往前冲==。\n\n它主要用于推理提速，常见于大模型生成优化。", "humanExplainDisplay": "以前是一颗一颗往外蹦字，\n现在干脆==一把吐几颗==，\n像输入法提前读懂你，\n直接==整段往前冲==。\n\n它主要用于推理提速，\n常见于大模型生成优化。", "relationsNarrative": "Token\n原本逐个 token 往外蹦，MTP 改成一次预测好几个。\n\nInference\nmulti-token prediction 主要发生在推理阶段，目标是让生成更快。\n\nLlm-inference-engine\nmulti-token prediction 往往需要推理引擎配合，才能稳定落地。\n\nContinuous batching\n连续批处理管调度，MTP 管一次吐几个，二者常一起提速。", "relations": { "token": { "label": "一次预测多个…", "note": "它把逐个生成改成成组生成。" }, "inference": { "label": "用于加速…", "note": "它属于推理阶段的速度优化手段。" }, "llm-inference-engine": { "label": "由…落地实现", "note": "这类能力通常靠推理引擎调度支持。" }, "continuous-batching": { "label": "可配合…提速", "note": "两者都常用于提升在线生成效率。" } } } } }, { "id": "multi-turn-medical-consultation", "name": "Multi-turn Medical Consultation", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-19T04:00:00.000Z", "relations": [ { "to": "ai-medical-assistant-ai" }, { "to": "agent-memory" }, { "to": "stateful-agent" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "Multi-turn Medical Consultation", "factExplain": "AI asks back-and-forth health questions over several chat turns to help judge symptoms.", "humanExplain": "It is like the school nurse with a clipboard. No magic ice pack yet, only more questions.\n\nYou meet it in online symptom checks and follow-up chats. It fills in health details, but a doctor still makes the call.", "humanExplainDisplay": "It is like the ==school nurse==\nwith a clipboard.\nNo ==magic ice pack== yet,\nonly more questions.\n\nYou meet it in online symptom checks\nand follow-up chats.\nIt fills in health details,\nbut a doctor still makes the call.", "relationsNarrative": "AI Medical Assistant\nMulti-turn consultation is the main chat style for many medical assistants.\n\nMemory\nIt uses Memory to remember earlier symptoms and health history.\n\nStateful agent\nA Stateful agent keeps the chat context across turns.\n\nHuman-in-the-loop\nHuman-in-the-loop lets a doctor review high-risk advice.", "relations": { "ai-medical-assistant-ai": { "label": "is the core chat for …", "note": "Medical assistants often ask follow-up questions to fill symptom gaps." }, "agent-memory": { "label": "uses … to remember history", "note": "Past symptoms and answers must stay in the chat." }, "stateful-agent": { "label": "is often built as …", "note": "The chat must keep context across turns." }, "human-in-the-loop": { "label": "needs … for backup", "note": "Doctors should review high-risk advice." } } }, "zh": { "fullName": "多轮医疗问诊", "factExplain": "AI 在多轮对话中持续追问并辅助判断病情。", "humanExplain": "像老中医搭脉后不急着下方子，先追着问寒热、吃睡、疼多久，线索越问越清楚。\n\n常见于线上分诊、慢病随访和问诊助手，能补全病情信息，但不能替医生拍板。", "humanExplainDisplay": "像老中医搭脉后，\n不急着==下方子==，\n先追问寒热吃睡，\n把线索==越问越清==。\n\n常见于线上分诊、\n慢病随访和问诊助手，\n能补全病情信息，\n但不能替医生拍板。", "relationsNarrative": "AI Medical Assistant\n多轮问诊是医疗助手最常见的核心交互方式。\n\nMemory\n它要记住前几轮症状和病史，追问才不会失忆。\n\nStateful agent\n问诊通常要保留上下文状态，才能连续追问。\n\nHuman-in-the-loop\n涉及高风险建议时，仍需要医生复核兜底。", "relations": { "ai-medical-assistant-ai": { "label": "是…的核心交互", "note": "医疗助手常靠连续追问补全病情。" }, "agent-memory": { "label": "依赖…记住病史", "note": "前文症状和回答要持续保留。" }, "stateful-agent": { "label": "常做成…", "note": "问诊需要跨轮保持上下文状态。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "高风险判断仍应由医生复核。" } } } } }, { "id": "multilingual-ai", "name": "Multilingual AI", "layer": "L4", "era": "2010s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "llm" }, { "to": "real-time-ai-translation" }, { "to": "transfer-learning" } ], "track": "today", "i18n": { "en": { "fullName": "Multilingual AI", "factExplain": "AI that can understand, write, or handle more than one language.", "humanExplain": "Multilingual AI is like an airport helper with magic ears. One traveler speaks Spanish. The next speaks Korean. It keeps smiling.\n\nIt translates text and chats with customers. It helps one product serve people in many languages.", "humanExplainDisplay": "Multilingual AI is like an ==airport helper==\nwith ==magic ears==.\nOne traveler speaks Spanish.\nThe next speaks Korean.\nIt keeps smiling.\n\nIt translates text and chats with customers.\nIt helps one product serve people\nin many languages.", "relationsNarrative": "NLP\nMultilingual AI is a core use of NLP.\n\nLLM\nMany multilingual AI systems use an LLM to learn several languages.\n\nReal-time AI Translation\nReal-time AI Translation is the most direct use of multilingual AI.\n\nTransfer Learning\nTransfer Learning helps move skills from one language to another.", "relations": { "natural-language-processing": { "label": "is part of …", "note": "Multilingual AI is a key use of NLP." }, "llm": { "label": "is often built with …", "note": "Many multilingual systems use LLMs to learn many languages." }, "real-time-ai-translation": { "label": "powers …", "note": "Real-time AI Translation is a common use of multilingual AI." }, "transfer-learning": { "label": "shares skills through …", "note": "Transfer Learning can move language skills from one language to another." } } }, "zh": { "fullName": "多语言 AI", "factExplain": "能理解、生成或处理多种语言的 AI 系统。", "humanExplain": "像景区里那个八国语言随时切换的金牌导游，中国游客、老外游客来一拨，它都能立刻接上话。\n\n常用于翻译、跨语种客服和全球产品，让一套系统服务多语言用户。", "humanExplainDisplay": "像景区里那个\n==八国语言随时切换==的\n==金牌导游==，\n中国游客、老外游客\n来一拨，\n它都能立刻接上话。\n\n常用于翻译、\n跨语种客服和全球产品，\n让一套系统服务\n多语言用户。", "relationsNarrative": "Natural-language-processing\n它是自然语言处理中的核心应用方向之一。\n\nLLM\n如今很多多语言能力，都是靠大模型统一学出来的。\n\nReal-time-ai-translation\n实时翻译是多语言能力最直接、最常见的落地场景。\n\nTransfer-learning\n它常利用迁移学习，把一种语言的能力带到另一种语言。", "relations": { "natural-language-processing": { "label": "属于…应用方向", "note": "它是自然语言处理的重要分支。" }, "llm": { "label": "常由…实现", "note": "很多多语言能力建立在大模型上。" }, "real-time-ai-translation": { "label": "支撑…落地", "note": "实时翻译是它最典型的应用。" }, "transfer-learning": { "label": "可借助…迁移", "note": "一种语言学到的能力可迁到另一种。" } } } } }, { "id": "multimodal-rag", "name": "MM RAG", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "rag" }, { "to": "multimodal" }, { "to": "vector-search" }, { "to": "document-parsing" } ], "track": "today", "i18n": { "en": { "fullName": "Multimodal Retrieval-Augmented Generation", "factExplain": "A RAG that searches images, tables, and audio before writing an answer.", "humanExplain": "MM RAG is the school librarian with superhero glasses and earbuds. Screenshots, tables, and voice notes do not scare it.\n\nIt looks through mixed files before it answers. It helps with contracts, reports, and security footage, but weak search can sink it.", "humanExplainDisplay": "MM RAG is the school librarian\nwith ==superhero glasses== and ==earbuds==.\nScreenshots, tables, and voice notes\ndo not scare it.\n\nIt looks through mixed files\nbefore it answers.\nIt helps with contracts, reports,\nand security footage.\nBut weak search can sink it.", "relationsNarrative": "RAG\nMM RAG upgrades RAG by adding images, audio, and video as sources.\n\nMultimodal AI\nMultimodal skills let MM RAG understand more than text.\n\nVector search\nMM RAG often uses Vector search to find related content in mixed files.\n\nDocument parsing\nDoc parsing organizes complex files so MM RAG can search them better.", "relations": { "rag": { "label": "extends … beyond text", "note": "It lets RAG search images, audio, video, and tables too." }, "multimodal": { "label": "brings … into search", "note": "Multimodal skills let images and audio be stored and searched." }, "vector-search": { "label": "finds content with …", "note": "Vector search helps it pull matching items from mixed media." }, "document-parsing": { "label": "depends on … to split files", "note": "Doc parsing cleans complex files before search." } } }, "zh": { "fullName": "多模态检索增强生成", "factExplain": "能检索图片、表格、音频等资料再生成回答的 RAG。", "humanExplain": "别人找资料只会翻文字，它更像望闻问切的老中医：截图、表格、录音都能搭脉，找料不靠一双眼睛死磕。\n\n适合合同、报告、监控这类图文音混杂的问答，信息更全，但很吃检索质量。", "humanExplainDisplay": "别人找资料只会翻文字，\n它更像==望闻问切的老中医==：\n截图、表格、录音都能搭脉，\n找料不靠==一双眼睛==死磕。\n\n适合合同、报告、监控这类\n图文音混杂的问答，\n信息更全，但很吃检索质量。", "relationsNarrative": "RAG\n它是 RAG 的升级版，把外部资料扩到图文音视频。\n\nMultimodal AI\n多模态能力让它不只看文字，也能理解图片音频。\n\nVector search\n它常靠向量搜索，从混合资料里找相关内容。\n\nDocument parsing\n复杂文档先被解析整理，它后面的检索才更靠谱。", "relations": { "rag": { "label": "扩展…输入源", "note": "它把检索对象从文本扩到多种模态。" }, "multimodal": { "label": "把…接入检索", "note": "多模态能力让图片音频也能入库查询。" }, "vector-search": { "label": "用…找内容", "note": "跨模态检索常靠向量搜索完成召回。" }, "document-parsing": { "label": "依赖…拆资料", "note": "复杂文档先解析，检索效果才更稳。" } } } } }, { "id": "multimodal", "name": "Multimodal AI", "layer": "L4", "era": "2023", "publishedAt": "2026-05-23T09:10:00Z", "relations": [ { "to": "llm" }, { "to": "embedding" }, { "to": "diffusion" }, { "to": "foundation-model" } ], "i18n": { "en": { "fullName": "Multimodal AI", "factExplain": "AI that can understand or create text, images, sound, and video.", "humanExplain": "Multimodal AI is like a kid with the TV sound back on. It sees the pie hit the face and hears the splat.\n\nIt lets AI read pictures and hear audio. You meet it when an app answers questions about a photo or a video.", "humanExplainDisplay": "Multimodal AI is like a kid with the ==TV sound back on==.\nIt sees the pie hit the face\nand hears the ==splat==.\n\nIt lets AI read pictures and hear audio.\nYou meet it when an app answers questions\nabout a photo or a video.", "relationsNarrative": "LLM\nMultimodal AI gives the LLM more than words to read and make.\n\nEmbedding\nEmbedding helps different media match up inside the AI.\n\nDiffusion\nDiffusion helps Multimodal AI make images.\n\nFoundation-model\nFoundation-model gives Multimodal AI one shared base.", "relations": { "llm": { "label": "expands … beyond text", "note": "Multimodal AI adds images and sound to the LLM's words." }, "embedding": { "label": "connects modes with …", "note": "Embeddings help AI compare words, pictures, and sound." }, "diffusion": { "label": "makes images with …", "note": "Diffusion gives Multimodal AI its image-making skill." }, "foundation-model": { "label": "upgrades … with more senses", "note": "A Foundation-model can share one base across many media types." } } }, "zh": { "fullName": "多模态", "factExplain": "AI 同时理解或生成文本、图像、语音等多种信息形式的能力。", "humanExplain": "多模态像终于不只会看字幕的 AI，图片、声音、文字都能一起听懂。\n\n它让模型能读图、听音、看视频，产品体验也从聊天框扩展到现实世界。", "humanExplainDisplay": "多模态像 AI 终于不只会看字幕，\n开始==看图、听声、读文字==。\n感官一下子多了起来。\n\n它让模型能读图、听音、看视频。\n聊天框里的 AI，\n开始往现实世界探头。", "relationsNarrative": "LLM\nLLM 为 Multimodal 系统提供语言理解和表达能力。\n\nEmbedding\nEmbedding 让不同模态进入可比较的表示空间。\n\nDiffusion\nDiffusion 扩展了 Multimodal 的图像生成能力。\n\nFoundation-model\nFoundation-model 让多模态能力共享同一个通用底座。", "relations": { "llm": { "label": "扩展…的输入输出" }, "embedding": { "label": "用…连接模态" }, "diffusion": { "label": "用…做视觉生成" }, "foundation-model": { "label": "让…升级" } } } } }, { "id": "multitask-learning", "name": "Multitask Learning", "layer": "L2", "era": "1997", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "transfer-learning" }, { "to": "representation-learning" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Multitask Learning", "factExplain": "One model learns several related tasks during the same training run.", "humanExplain": "Multitask learning is like learning to juggle while riding a unicycle. Your balance gets better at both, even though they look different.\n\nRelated tasks train together and share useful patterns. You meet it in voice AI, image AI, and writing tools.", "humanExplainDisplay": "Multitask learning is like ==learning to juggle==\nwhile riding a unicycle.\nYour ==balance== gets better at both,\neven though they look different.\n\nRelated tasks train together\nand share useful patterns.\nYou meet it in voice AI,\nimage AI,\nand writing tools.", "relationsNarrative": "Transfer Learning\nMultitask Learning helps related tasks share useful lessons.\n\nRepresentation Learning\nSeveral tasks push the model to learn one shared inner picture.\n\nSupervised Learning\nIt often puts several labeled tasks into the same training run.", "relations": { "transfer-learning": { "label": "helps … happen", "note": "Multitask training can move useful skills between related tasks." }, "representation-learning": { "label": "shares …", "note": "Several tasks push the model to learn a general inner picture." }, "supervised-learning": { "label": "extends …", "note": "It often trains several labeled tasks in one run." } } }, "zh": { "fullName": "多任务学习", "factExplain": "一种让模型同时学习多个相关任务的训练方法。", "humanExplain": "多任务学习像餐厅后厨轮岗：切菜的刀工帮了炒菜，炒菜的火候又帮了炖汤。\n\n用于语音、视觉、文本，让模型学出一个共用的底层理解，而不是只会单项。", "humanExplainDisplay": "多任务学习像\n餐厅后厨轮岗：\n==切菜的刀工帮了炒菜==，\n炒菜的火候\n又帮了炖汤。\n\n用于语音、视觉、文本，\n让模型学出一个\n共用的底层理解，\n而不是只会单项。", "relationsNarrative": "Transfer Learning\n多任务学习让相关任务之间更容易迁移经验。\n\nRepresentation Learning\n多个任务共享底层表示，减少只会单项的偏科。\n\nSupervised Learning\n它常把多个有标签任务放进同一训练流程。", "relations": { "transfer-learning": { "label": "促进…", "note": "多任务训练常带来任务间迁移。" }, "representation-learning": { "label": "共享…", "note": "多个任务逼模型学通用表示。" }, "supervised-learning": { "label": "扩展…", "note": "常把多个有标签任务合并训练。" } } } } }, { "id": "mycin", "name": "MYCIN", "layer": "L4", "era": "1972", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "expert-system" }, { "to": "production-system" }, { "to": "symbolic-ai" }, { "to": "dendral" } ], "track": "history", "i18n": { "en": { "fullName": "MYCIN", "factExplain": "A Stanford rule-based expert system for medical diagnosis.", "humanExplain": "MYCIN was like a doctor with a giant checklist. No small talk, just infection advice.\n\nIt turned expert rules into diagnosis tips. Doctors used it for infection drug advice.", "humanExplainDisplay": "MYCIN was like a doctor\nwith a ==giant checklist==.\nNo small talk,\njust ==infection advice==.\n\nIt turned expert rules\ninto diagnosis tips.\nDoctors used it\nfor infection drug advice.", "relationsNarrative": "Expert System\nMYCIN is the classic medical case of an Expert System.\n\nProduction\nMYCIN used rules to move from symptoms to a diagnosis.\n\nSymbolic AI\nMYCIN followed the Symbolic AI path of rule-based reasoning.\n\nDENDRAL\nDENDRAL first showed expert knowledge could be written as rules.", "relations": { "expert-system": { "label": "is a kind of …", "note": "MYCIN is a classic medical Expert System." }, "production-system": { "label": "uses … rules", "note": "Production rules helped MYCIN move from symptoms to advice." }, "symbolic-ai": { "label": "follows …", "note": "MYCIN reasoned with symbols and written rules." }, "dendral": { "label": "continues …", "note": "DENDRAL showed expert knowledge could be written as rules." } } }, "zh": { "fullName": "早期医疗专家系统", "factExplain": "斯坦福开发的规则式医疗诊断专家系统。", "humanExplain": "MYCIN 就是老中医把脉口诀本：症状一来，照规则翻到感染判断。\n\n它把专家规则变成诊断建议，用在感染用药咨询场景。", "humanExplainDisplay": "MYCIN 就是\n==老中医把脉口诀本==：\n症状一来，\n照规则翻到感染判断。\n\n它把专家规则变成诊断建议，\n用在感染用药咨询场景。", "relationsNarrative": "Expert System\nMYCIN 是专家系统最经典的医疗案例。\n\nProduction\n它用规则库把症状一步步推向判断。\n\nSymbolic AI\nMYCIN 体现了用符号规则推理的路线。\n\nDENDRAL\nDENDRAL 先证明专家知识能写成规则。", "relations": { "expert-system": { "label": "属于…", "note": "MYCIN 是医疗专家系统代表。" }, "production-system": { "label": "用…写规则", "note": "规则让诊断像按流程查表。" }, "symbolic-ai": { "label": "体现…路线", "note": "它依赖符号和规则推理。" }, "dendral": { "label": "延续…思路", "note": "二者同属早期规则派代表。" } } } } }, { "id": "n-gram-language-model", "name": "N-gram LM", "layer": "L2", "era": "1980", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "language-modeling" }, { "to": "autoregressive-model" }, { "to": "token" }, { "to": "perplexity" } ], "track": "history", "i18n": { "en": { "fullName": "n-gram Language Model", "factExplain": "A stats-based language model that guesses the next token from the previous n−1 tokens.", "humanExplain": "An n-gram LM is like your phone keyboard. It peeks at a few words, then proudly suggests “sounds good.”\n\nIt helped keyboards and speech tools guess text. Long stories can confuse it.", "humanExplainDisplay": "An n-gram LM is like ==your phone keyboard==.\nIt peeks at ==a few words==,\nthen proudly suggests “sounds good.”\n\nIt helped keyboards and speech tools guess text.\nLong stories can confuse it.", "relationsNarrative": "LM\nAn n-gram LM was a classic early stats-based way to do LM.\n\nAutoregressive Model\nIt predicts the next token from left to right.\n\nToken\nTokens are the basic pieces it counts and predicts.\n\nPerplexity\nPerplexity is often used to test how well it guesses the next token.", "relations": { "language-modeling": { "label": "was an early kind of …", "note": "It is a classic stats-based way to do LM." }, "autoregressive-model": { "label": "is a kind of …", "note": "It predicts the next token from earlier tokens, in order." }, "token": { "label": "counts … sequences", "note": "Tokens are the pieces it counts and guesses with." }, "perplexity": { "label": "is tested with …", "note": "Perplexity scores how well it guesses the next token." } } }, "zh": { "fullName": "n 元语言模型", "factExplain": "用前 n-1 个词预测下一个词的统计语言模型。", "humanExplain": "n-gram像输入法顺嘴接话：只瞄前几个词，就敢猜你下一句是“收到”。\n\n早年撑起手机联想和语音识别，长上下文容易迷糊。", "humanExplainDisplay": "n-gram像输入法顺嘴接话：\n==只瞄前几个词==，\n就敢猜你下一句是\n==“收到”==。\n\n早年撑起手机联想\n和语音识别，\n长上下文容易迷糊。", "relationsNarrative": "Language Modeling\n它是语言建模早期最经典的统计方法。\n\nAutoregressive Model\n它按从左到右的顺序预测下一个词元。\n\nToken\n词元是它统计 n 连串和预测的基本单位。\n\nPerplexity\n困惑度常用来评估它猜下个词准不准。", "relations": { "language-modeling": { "label": "实现早期…", "note": "它是统计语言建模的经典路线。" }, "autoregressive-model": { "label": "属于…", "note": "它按顺序用前文预测下个词。" }, "token": { "label": "统计…序列", "note": "词元是它计数和预测的基本单位。" }, "perplexity": { "label": "用…评估", "note": "困惑度常用来衡量猜词好坏。" } } } } }, { "id": "naive-bayes", "name": "Naive Bayes", "layer": "L2", "era": "1960", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "bayesian-network" }, { "to": "bag-of-words" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Naive Bayes", "factExplain": "A classification method that treats each feature as separate when choosing a class.", "humanExplain": "Naive Bayes is like a spam bouncer at a club. It counts each sketchy word, then pretends those words never met.\n\nYou meet it in spam filters and text sorting. It is fast and cheap, but linked clues can fool it.", "humanExplainDisplay": "Naive Bayes is like a ==spam bouncer== at a club.\nIt ==counts each sketchy word==,\nthen pretends those words never met.\n\nYou meet it in spam filters and text sorting.\nIt is fast and cheap,\nbut linked clues can fool it.", "relationsNarrative": "Classification\nNaive Bayes does Classification by choosing the most likely class.\n\nBayesian Network\nNaive Bayes is a very simple Bayesian Network with a strong independence rule.\n\nBag-of-Words\nBag-of-Words turns text into word-count features for Naive Bayes.\n\nSupervised Learning\nNaive Bayes uses Supervised Learning with labeled examples.", "relations": { "classification": { "label": "used for …", "note": "It outputs the class an item most likely belongs to." }, "bayesian-network": { "label": "simplifies …", "note": "A strong independence rule makes the structure very simple." }, "bag-of-words": { "label": "often uses …", "note": "Word counts become features it can classify with." }, "supervised-learning": { "label": "is part of …", "note": "It learns from examples with labels." } } }, "zh": { "fullName": "朴素贝叶斯", "factExplain": "一种基于特征条件独立假设的分类算法。", "humanExplain": "朴素贝叶斯像门口安检：每个可疑词单独记一笔，然后假装这些词从没串过供。\n\n常做垃圾邮件和文本分类；快省，相关特征会误导。", "humanExplainDisplay": "朴素贝叶斯像门口安检：\n每个可疑词单独记一笔，\n然后假装这些词==从没串过供==。\n\n常做垃圾邮件和文本分类；\n快省，\n相关特征会误导。", "relationsNarrative": "Classification\nNaive Bayes 是常见分类器，输出最可能的类别。\n\nBayesian Network\n它可看作结构极简、假设很强的贝叶斯网络。\n\nBag-of-Words\n词袋把文本拆成特征，方便它做垃圾邮件分类。\n\nSupervised Learning\n它用带标签数据学习各类别的概率。", "relations": { "classification": { "label": "用于…", "note": "它输出样本最可能所属的类别。" }, "bayesian-network": { "label": "简化自…", "note": "强独立假设让结构极简。" }, "bag-of-words": { "label": "常搭配…", "note": "文本词频可直接变成分类特征。" }, "supervised-learning": { "label": "属于…", "note": "训练时需要已标注样本。" } } } } }, { "id": "named-entity-recognition", "name": "NER", "layer": "L4", "era": "1995", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "information-extraction" }, { "to": "sequence-labeling" }, { "to": "natural-language-processing" }, { "to": "knowledge-graph" } ], "track": "history", "i18n": { "en": { "fullName": "Named Entity Recognition", "factExplain": "NER finds and labels names, places, companies, and other key things in text.", "humanExplain": "NER is the school office sorting lost-and-found notes. It circles “Mia,” “Room 12,” and “Chess Club” before the hoodie goes on a world tour.\n\nIt pulls key names and places from text. It helps content checks. It also helps Knowledge Graphs and tracking online talk.", "humanExplainDisplay": "NER is the ==school office== sorting lost-and-found notes.\nIt ==circles “Mia,” “Room 12,” and “Chess Club”==\nbefore the hoodie goes on a world tour.\n\nIt pulls key names and places from text.\nIt helps content checks.\nIt also helps Knowledge Graphs\nand tracking online talk.", "relationsNarrative": "IE\nNER gives IE entities before it finds relations.\n\nSequence Labeling\nNER often uses Sequence Labeling to tag each word.\n\nNLP\nNER is a classic NLP task for understanding text.\n\nKnowledge Graph\nNER gives Knowledge Graphs the nodes they need.", "relations": { "information-extraction": { "label": "is a base task for …", "note": "Finding entities first makes later relation extraction steadier." }, "sequence-labeling": { "label": "often uses …", "note": "It tags each word as an entity type or not." }, "natural-language-processing": { "label": "belongs to …", "note": "NER is a classic text understanding task in NLP." }, "knowledge-graph": { "label": "extracts nodes for …", "note": "Found entities often become nodes in the graph." } } }, "zh": { "fullName": "命名实体识别", "factExplain": "从文本中识别并分类人名、地名等实体。", "humanExplain": "NER 像快递员读面单：姓名、地址、公司先圈出，信息包裹才不会满城乱飞。\n\n用于抽取文本关键实体，服务审查、图谱和舆情。", "humanExplainDisplay": "NER 像快递员读面单：\n姓名、地址、公司先==圈出==，\n信息包裹才不会==满城乱飞==。\n\n用于抽取文本关键实体，\n服务审查、\n图谱和舆情。", "relationsNarrative": "Information Extraction\n命名实体识别先找出实体，为后续关系抽取打底。\n\nSequence Labeling\n它常被做成序列标注任务，逐词判断实体类型。\n\nNLP\n它是自然语言处理中经典的文本理解任务。\n\nKnowledge Graph\n识别出的实体常会变成知识图谱里的节点。", "relations": { "information-extraction": { "label": "是…的基础任务", "note": "先找实体，后续抽关系更稳。" }, "sequence-labeling": { "label": "常用…来实现", "note": "把每个词标成实体或非实体。" }, "natural-language-processing": { "label": "属于…任务", "note": "它是文本理解里的经典任务。" }, "knowledge-graph": { "label": "为…抽取节点", "note": "实体常变成图谱里的节点。" } } } } }, { "id": "national-security-ai", "name": "National security AI", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "sovereign-ai" }, { "to": "ai-export-controls" }, { "to": "ai-war-game-simulation" }, { "to": "ai-governance-framework" } ], "track": "today", "i18n": { "en": { "fullName": "National Security AI", "factExplain": "AI systems and rules used for national security work.", "humanExplain": "National security AI is like a country-sized smoke alarm. Great for real danger. Annoying if it screams over burnt toast.\n\nIt helps intelligence teams, cyber defense crews, and battlefield planners. The big questions are access and blame.", "humanExplainDisplay": "National security AI is like\na ==country-sized smoke alarm==.\nGreat for real danger.\nAnnoying if it screams over ==burnt toast==.\n\nIt helps intelligence teams,\ncyber defense crews,\nand battlefield planners.\nThe big questions are access and blame.", "relationsNarrative": "Sovereign AI\nNational Security AI often depends on local computing power and models a country can control.\n\nAI Export Controls\nAI Export Controls put key chips and models inside a security boundary.\n\nAI Wargaming Simulation\nAI Wargaming Simulation is a common use in defense planning.\n\nAI Governance\nThe higher the risk, the more National Security AI needs clear governance rules.", "relations": { "sovereign-ai": { "label": "depends on …", "note": "National security needs AI power a country can control." }, "ai-export-controls": { "label": "triggers …", "note": "Key chips and models often fall under export controls." }, "ai-war-game-simulation": { "label": "used for …", "note": "Wargaming is a common defense use case." }, "ai-governance-framework": { "label": "needs rules from …", "note": "High-risk security work needs clear rules and backup." } } }, "zh": { "fullName": "国家安全 AI", "factExplain": "用于国家安全任务的 AI 系统与治理实践。", "humanExplain": "国家安全AI像给棋盘装了天眼：能提前看杀招，也必须防误判和越界。\n\n用于情报、网防、战场决策，核心是权限与问责。", "humanExplainDisplay": "国家安全AI像给棋盘\n装了==天眼==：\n能提前看杀招，\n也必须防==误判和越界==。\n\n用于情报、网防、战场决策，\n核心是权限与问责。", "relationsNarrative": "Sovereign AI\n国家安全 AI 往往依赖本土可控的算力与模型。\n\nAI Export Controls\n出口管制会把关键芯片和模型纳入安全边界。\n\nAI Wargaming Simulation\n兵棋推演是它在国防决策中的典型用法。\n\nAI Governance\n安全场景越高风险，越需要治理框架约束。", "relations": { "sovereign-ai": { "label": "依赖…能力", "note": "国家安全需要本土可控的 AI 能力。" }, "ai-export-controls": { "label": "触发…", "note": "关键芯片和模型常被纳入管制。" }, "ai-war-game-simulation": { "label": "用于…", "note": "兵棋推演是典型国防应用场景。" }, "ai-governance-framework": { "label": "需要…约束", "note": "高风险场景更需要规则兜底。" } } } } }, { "id": "natural-language-processing", "name": "NLP", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "language-modeling" }, { "to": "bert" }, { "to": "llm" }, { "to": "information-retrieval" } ], "track": "history", "i18n": { "en": { "fullName": "Natural Language Processing", "factExplain": "A field for making computers understand, process, and create human language.", "humanExplain": "NLP is a computer in a family group chat. It learns “fine” can mean “fine” or “you are in trouble.”\n\nYou meet it in chatbots and translation apps. It also helps search boxes and support chats.", "humanExplainDisplay": "NLP is a computer in a ==family group chat==.\nIt learns ==“fine” can mean “fine” or “you are in trouble.”==\n\nYou meet it in chatbots and translation apps.\nIt also helps search boxes and support chats.", "relationsNarrative": "LM\nLanguage modeling is one of the core methods in NLP.\n\nBERT\nBERT is one of the landmark models from the NLP era.\n\nLLM\nLLMs build on many years of NLP work.\n\nIR\nNLP often works with IR for search, Q&A, and knowledge services.", "relations": { "language-modeling": { "label": "gave rise to …", "note": "Language modeling is a core path inside NLP." }, "bert": { "label": "produced …", "note": "BERT is a classic milestone in NLP." }, "llm": { "label": "laid groundwork for …", "note": "LLMs build on years of NLP work." }, "information-retrieval": { "label": "often works with …", "note": "NLP and IR often team up for search and Q&A." } } }, "zh": { "fullName": "自然语言处理（Natural Language Processing）", "factExplain": "让计算机理解、处理和生成人类语言的技术领域。", "humanExplain": "自然语言处理像相亲时会读空气：不只听你说了啥，还得听懂潜台词，别把场子聊冷。\n\n它支撑聊天、翻译、搜索和客服，是机器处理人类语言的核心技术。", "humanExplainDisplay": "自然语言处理像相亲时会==读空气==：\n不只听你说了啥，\n还得听懂==潜台词==，\n别把场子聊冷。\n\n它支撑聊天、\n翻译、\n搜索和客服，\n是机器处理人类语言的核心技术。", "relationsNarrative": "Language Modeling\n语言建模是自然语言处理里的核心方法路线之一。\n\nBERT\nBERT 是自然语言处理时代的代表性模型之一。\n\nLLM\n大语言模型建立在自然语言处理的长期积累之上。\n\nInformation Retrieval\n它常和信息检索结合，做搜索、问答与知识服务。", "relations": { "language-modeling": { "label": "催生…方法", "note": "它的重要路线之一就是语言建模。" }, "bert": { "label": "产出…代表模型", "note": "BERT 是它经典里程碑之一。" }, "llm": { "label": "为…打基础", "note": "大语言模型建立在它长期积累上。" }, "information-retrieval": { "label": "常与…结合", "note": "搜索与问答系统里经常一起用。" } } } } }, { "id": "natural-language-understanding", "name": "NLU", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "llm" }, { "to": "classification" }, { "to": "information-extraction" } ], "track": "history", "i18n": { "en": { "fullName": "Natural Language Understanding", "factExplain": "A technique that helps machines understand text meaning and intent.", "humanExplain": "NLU is your friend reading a group chat. It sees “fine” and spots the storm cloud you did not send.\n\nIt helps customer support, search, and assistants. It reads the need before choosing the action.", "humanExplainDisplay": "NLU is your friend\nreading a ==group chat==.\nIt sees “fine”\nand spots the ==storm cloud==\nyou did not send.\n\nIt helps customer support,\nsearch,\nand assistants.\nIt reads the need\nbefore choosing the action.", "relationsNarrative": "NLP\nNLU is the part of NLP that reads meaning and intent.\n\nLLM\nAn LLM uses language understanding to make conversation work.\n\nClassification\nIntent detection is often modeled as a classification problem.\n\nIE\nIE needs text understanding to find entities and relations.", "relations": { "natural-language-processing": { "label": "is part of …", "note": "NLU is the part of NLP that reads meaning and intent." }, "llm": { "label": "supports … chat", "note": "LLMs need language understanding to hold useful conversations." }, "classification": { "label": "uses … for intent", "note": "Intent detection is often treated as a classification task." }, "information-extraction": { "label": "helps … find meaning", "note": "IE needs understanding to find entities and relations in text." } } }, "zh": { "fullName": "自然语言理解", "factExplain": "让机器识别文本含义与意图的技术。", "humanExplain": "NLU像听懂领导说你看着办：字面很客气，真正要你补方案背锅。\n\n它用于客服、搜索和助手，先懂需求再决定动作。", "humanExplainDisplay": "NLU像听懂领导说\n==你看着办==：\n字面很客气，\n真正要你==补方案背锅==。\n\n它用于客服、搜索和助手，\n先懂需求，\n再决定动作。", "relationsNarrative": "NLP\n自然语言理解是 NLP 中负责读懂语义和意图的部分。\n\nLLM\nLLM 的对话能力，很大程度依赖语言理解能力。\n\nClassification\n意图识别常被建模成分类问题。\n\nInformation Extraction\n信息抽取需要先理解文本里的实体和关系。", "relations": { "natural-language-processing": { "label": "属于…", "note": "理解是 NLP 的核心分支。" }, "llm": { "label": "支撑…对话", "note": "大模型好用，离不开理解能力。" }, "classification": { "label": "常用…识别意图", "note": "意图识别常被做成分类任务。" }, "information-extraction": { "label": "帮助…抽含义", "note": "抽实体和关系，需要语义理解。" } } } } }, { "id": "needle-in-a-haystack-test", "name": "Needle-haystack", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "context-window" }, { "to": "rag" }, { "to": "benchmark-contamination" } ], "track": "today", "i18n": { "en": { "fullName": "Needle-in-a-Haystack Test", "factExplain": "A test of whether a model can find one fact in a long input.", "humanExplain": "Needle-haystack is Where’s Waldo for AI, but the page is a phone book. One tiny sentence hides in a giant stack.\n\nPeople use it on long documents and code. It shows if the model forgets facts as the text gets longer.", "humanExplainDisplay": "Needle-haystack is ==Where’s Waldo for AI==,\nbut the page is a ==phone book==.\nOne tiny sentence hides in a giant stack.\n\nPeople use it on long documents and code.\nIt shows if the model forgets facts\nas the text gets longer.", "relationsNarrative": "Context-window\nNeedle-haystack tests if the model can find facts in a long context.\n\nRAG\nNeedle-haystack tests finding a needle in the input, not outside lookup.\n\nBenchmark contamination\nIf test samples leak, the score may look good but not be trustworthy.", "relations": { "context-window": { "label": "tests the limit of …", "note": "Longer context needs proof the model can still find faraway facts." }, "rag": { "label": "adds a check beside …", "note": "It tests search inside the input, not outside lookup." }, "benchmark-contamination": { "label": "can be fooled by …", "note": "Leaked test items can make scores look too high." } } }, "zh": { "fullName": "大海捞针测试", "factExplain": "一种检验模型长上下文检索能力的测试。", "humanExplain": "这测试像班主任把考点塞进八十页讲义，翻到最后突然抽问，看你还能不能把那句小字原封不动找出来。\n\n常拿来测长文问答、文档分析、代码排查时，模型会不会越看越失忆。", "humanExplainDisplay": "这测试像班主任把考点\n塞进==八十页讲义==，\n翻到最后突然抽问，\n看你还能不能==原封找出==。\n\n常拿来测长文问答、\n文档分析、代码排查时，\n模型会不会\n越看越失忆。", "relationsNarrative": "Context-window\n它常被用来检验长上下文里找信息的真实能力。\n\nRAG\n它测模型在长输入里找针，不等于测外部检索能力。\n\nBenchmark contamination\n若测试样本泄露，成绩可能好看却不可信。", "relations": { "context-window": { "label": "检验…上限", "note": "上下文越长，越需要测远处信息能否找回。" }, "rag": { "label": "补充评估…", "note": "它测记忆检索，RAG测外部资料调用。" }, "benchmark-contamination": { "label": "防止…误导", "note": "题目泄露会让测试成绩虚高失真。" } } } } }, { "id": "neocognitron", "name": "Neocognitron", "layer": "L3", "era": "1980", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "lenet-5" }, { "to": "neural-network" }, { "to": "computer-vision" } ], "track": "history", "i18n": { "en": { "fullName": "Neocognitron", "factExplain": "An early layered neural network for recognizing patterns in images.", "humanExplain": "Neocognitron is like spotting a cat in a messy sticker book. First whiskers, then ears, then: yep, cat.\n\nIt came before CNNs. It helped image AI build small clues into bigger shapes.", "humanExplainDisplay": "Neocognitron is like spotting a cat\nin a ==messy sticker book==.\nFirst ==whiskers==,\nthen ears,\nthen: yep, cat.\n\nIt came before CNNs.\nIt helped image AI build small clues\ninto bigger shapes.", "relationsNarrative": "CNN\nNeocognitron was an early ancestor of CNN.\n\nLeNet-5\nLeNet-5 followed its layered path for image recognition.\n\nNeural-network\nNeocognitron was an early neural network inspired by vision.\n\nComputer Vision\nNeocognitron aimed to help machines spot patterns in pictures.", "relations": { "cnn": { "label": "inspired …", "note": "CNNs kept its idea of small patches first, then larger shapes." }, "lenet-5": { "label": "paved the way for …", "note": "LeNet-5 followed the same layered image-recognition path." }, "neural-network": { "label": "is a kind of …", "note": "It was an early neural network for vision." }, "computer-vision": { "label": "was used for …", "note": "Its goal was to help machines spot patterns in pictures." } } }, "zh": { "fullName": "新认知机", "factExplain": "一种早期分层视觉识别神经网络。", "humanExplain": "新认知机像拼乐高认猫：先找耳朵胡须，再拼整只猫才点头。\n\n它是卷积网络前辈，启发图像识别里的分层特征。", "humanExplainDisplay": "新认知机像\n==拼乐高认猫==：\n先找耳朵胡须，\n再拼整只猫才点头。\n\n它是卷积网络前辈，\n启发图像识别里的\n分层特征。", "relationsNarrative": "CNN\n新认知机是 CNN 的早期直系先祖。\n\nLeNet-5\nLeNet-5 延续了图像分层识别路线。\n\nNeural Network\n它是早期受视觉皮层启发的神经网络。\n\nComputer Vision\n它的目标是让机器识别图像模式。", "relations": { "cnn": { "label": "启发…", "note": "CNN 继承了分层局部特征思想。" }, "lenet-5": { "label": "铺路…", "note": "LeNet-5 延续了图像分层识别路线。" }, "neural-network": { "label": "属于…", "note": "它是早期视觉神经网络模型。" }, "computer-vision": { "label": "用于…", "note": "目标是让机器识别图像模式。" } } } } }, { "id": "nesterov-accelerated-gradient", "name": "NAG", "layer": "L2", "era": "1983", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "momentum" }, { "to": "sgd" }, { "to": "gradient-descent" }, { "to": "adam" } ], "track": "history", "i18n": { "en": { "fullName": "Nesterov Accelerated Gradient", "factExplain": "A momentum optimizer with a gradient check at the look-ahead point.", "humanExplain": "NAG is a back-seat driver, but useful for once. It looks down the road first. Then it yells “turn now!” before the car drifts.\n\nIt speeds up Gradient Descent in model training. It often acts as a smarter Momentum step for SGD.", "humanExplainDisplay": "NAG is a ==back-seat driver==,\nbut useful for once.\nIt ==looks down the road== first.\nThen it yells “turn now!”\nbefore the car drifts.\n\nIt speeds up Gradient Descent\nin model training.\nIt often acts as\na smarter Momentum step for SGD.", "relationsNarrative": "Momentum\nNAG is Momentum with a look-ahead check before the correction.\n\nSGD\nNAG often works as a faster update rule for SGD.\n\nGradient Descent\nNAG improves Gradient Descent so it can reach low loss faster.\n\nAdam\nAdam also uses momentum, but it changes the learning rate as it learns.", "relations": { "momentum": { "label": "adds look-ahead to …", "note": "NAG tries the momentum step first, then corrects with the gradient." }, "sgd": { "label": "speeds up …", "note": "NAG is often used as a faster update rule for SGD." }, "gradient-descent": { "label": "improves …", "note": "NAG helps the steps reach low loss faster." }, "adam": { "label": "shares momentum with …", "note": "Adam also uses momentum, and it changes the learning rate as it goes." } } }, "zh": { "fullName": "Nesterov 加速梯度", "factExplain": "先在前瞻位置计算梯度的动量优化算法。", "humanExplain": "NAG像老司机过弯：先看弯后路况再打方向，不等漂出去才补救。\n\n用于加速梯度下降，比普通动量更会提前修正。", "humanExplainDisplay": "NAG像老司机过弯：\n==先看弯后路况==\n再打方向，\n不等漂出去才补救。\n\n用于加速梯度下降，\n比普通动量\n更会提前修正。", "relationsNarrative": "Momentum\nNAG 是动量法的前瞻版，先预判再修正。\n\nSGD\n它常作为 SGD 的加速更新规则使用。\n\nGradient Descent\n它改造梯度下降，让收敛更快更稳。\n\nAdam\nAdam 也用动量思想，但会自适应调学习率。", "relations": { "momentum": { "label": "前瞻改造…", "note": "先按惯性探路，再用梯度修正。" }, "sgd": { "label": "加速…", "note": "常作为 SGD 的动量升级版。" }, "gradient-descent": { "label": "改进…", "note": "目标是更快靠近损失低点。" }, "adam": { "label": "共享动量思想", "note": "Adam 叠加了自适应学习率。" } } } } }, { "id": "network-pruning", "name": "Network Pruning", "layer": "L2", "era": "1989", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "model-compression" }, { "to": "quantization" }, { "to": "parameter" }, { "to": "inference" } ], "track": "history", "i18n": { "en": { "fullName": "Network Pruning", "factExplain": "A way to shrink a model by removing extra links or neurons.", "humanExplain": "Network pruning is like trimming a hedge before it swallows the mailbox. Snip the extra twigs, and the hedge still looks fine.\n\nIt removes weak links or neurons for Model Compression and phone-sized AI. It saves compute, but cut too much and scores drop.", "humanExplainDisplay": "Network pruning is like trimming a hedge\nbefore it ==swallows the mailbox==.\nSnip the ==extra twigs==,\nand the hedge still looks fine.\n\nIt removes weak links or neurons\nfor Model Compression and phone-sized AI.\nIt saves compute,\nbut cut too much and scores drop.", "relationsNarrative": "Model Compression\nNetwork pruning is a common Model Compression method. It directly makes the network smaller.\n\nQuantization\nQuantization changes number precision. Pruning changes the network structure.\n\nParameter\nPruning often checks Parameter importance and removes weak weights.\n\nInference\nAfter extra work is removed, Inference is often faster and lighter.", "relations": { "model-compression": { "label": "works as …", "note": "Pruning removes extra links so the model gets smaller." }, "quantization": { "label": "often pairs with …", "note": "Pruning cuts structure. Quantization lowers number precision." }, "parameter": { "label": "removes weak …", "note": "Pruning often cuts weights with low importance." }, "inference": { "label": "speeds up …", "note": "With less useless work, inference becomes lighter." } } }, "zh": { "fullName": "网络剪枝", "factExplain": "删除冗余连接或神经元以压缩模型的方法。", "humanExplain": "网络剪枝像给月季修枝：枯弱枝下剪，花架清爽，照样开得旺。\n\n用于模型压缩和端侧部署，省算力，剪狠会掉分。", "humanExplainDisplay": "网络剪枝像\n给月季修枝：\n==枯弱枝下剪==，\n花架清爽，照样开得旺。\n\n用于模型压缩和端侧部署，\n省算力，\n剪狠会掉分。", "relationsNarrative": "Model Compression\n网络剪枝是模型压缩的典型做法，直接缩小网络。\n\nQuantization\n量化改数字精度，剪枝改网络结构，常一起用。\n\nParameter\n剪枝常按参数重要性，删掉贡献小的权重。\n\nInference\n删掉冗余计算后，推理通常更快更省。", "relations": { "model-compression": { "label": "作为…手段", "note": "剪掉冗余连接，让模型变小。" }, "quantization": { "label": "常与…搭配", "note": "一个减结构，一个降精度。" }, "parameter": { "label": "删除冗余…", "note": "常按权重重要性下刀。" }, "inference": { "label": "加速…过程", "note": "少算无用部分，推理更轻。" } } } } }, { "id": "neural-architecture-search", "name": "NAS", "layer": "L2", "era": "2016", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "hyperparameter-optimization" }, { "to": "reinforcement-learning" }, { "to": "bayesian-optimization" } ], "track": "history", "i18n": { "en": { "fullName": "Neural Architecture Search", "factExplain": "A method that automatically searches for a good neural network structure.", "humanExplain": "NAS is a bake-off for neural nets. The recipe with the best cake stays.\n\nIt designs model structure by testing many options. This saves hand tuning, but the training bill can get scary.", "humanExplainDisplay": "NAS is a ==bake-off== for neural nets.\nThe ==recipe with the best cake== stays.\n\nIt designs model structure\nby testing many options.\nThis saves hand tuning,\nbut the training bill can get scary.", "relationsNarrative": "Neural-network\nNAS searches the structure of a neural network.\n\nHPO\nNAS can be seen as HPO with a bigger search space.\n\nRL\nEarly NAS often used RL to suggest candidate structures.\n\nBayesian Optimization\nBayesian Optimization helps NAS avoid costly trial and error.", "relations": { "neural-network": { "label": "searches … structure", "note": "NAS looks for a better neural network structure." }, "hyperparameter-optimization": { "label": "fits under …", "note": "NAS can treat architecture as a thing to tune." }, "reinforcement-learning": { "label": "uses … for candidates", "note": "Early NAS often used RL to suggest new structures." }, "bayesian-optimization": { "label": "uses … to cut trial and error", "note": "Bayesian Optimization helps avoid wasteful training runs." } } }, "zh": { "fullName": "神经架构搜索", "factExplain": "自动搜索神经网络结构的方法。", "humanExplain": "NAS像开武林擂台选门派：别吹祖传秘籍，谁实战能打就留下。\n\n自动设计模型结构，少靠人工试错，但训练很烧钱。", "humanExplainDisplay": "NAS像开武林擂台选门派：\n别吹==祖传秘籍==，\n谁==实战能打==就留下。\n\n自动设计模型结构，\n少靠人工试错，\n但训练很烧钱。", "relationsNarrative": "Neural Network\nNAS 搜索的对象，就是神经网络的结构。\n\nHyperparameter Optimization\nNAS 可看作搜索范围更大的超参数优化。\n\nReinforcement Learning\n早期 NAS 常用强化学习生成候选结构。\n\nBayesian Optimization\n贝叶斯优化常用来减少昂贵试错。", "relations": { "neural-network": { "label": "搜索…结构", "note": "目标是挑出更合适的网络结构。" }, "hyperparameter-optimization": { "label": "属于…范畴", "note": "架构也可被当作超参数搜索。" }, "reinforcement-learning": { "label": "用…生成候选", "note": "早期方法常让策略模型提方案。" }, "bayesian-optimization": { "label": "用…减少试错", "note": "适合在昂贵训练中少走弯路。" } } } } }, { "id": "neural-machine-translation", "name": "NMT", "layer": "L4", "era": "2014", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "seq2seq" }, { "to": "bahdanau-attention" }, { "to": "transformer" }, { "to": "real-time-ai-translation" } ], "track": "history", "i18n": { "en": { "fullName": "Neural Machine Translation", "factExplain": "A way to translate text between languages using neural networks.", "humanExplain": "NMT is the bilingual friend at the cafeteria table. It hears the whole sentence, then says it in normal human words.\n\nYou meet it on webpages and subtitles. You also see it in support chats. Tricky terms and rare languages still need a human check.", "humanExplainDisplay": "NMT is the ==bilingual friend==\nat the cafeteria table.\nIt hears the ==whole sentence==,\nthen says it in normal human words.\n\nYou meet it on webpages and subtitles.\nYou also see it in support chats.\nTricky terms and rare languages\nstill need a human check.", "relationsNarrative": "Seq2Seq\nSeq2Seq was a main early setup for neural translation.\n\nBahdanau Attention\nBahdanau Attention helps the model match key words in long sentences.\n\nTransformer\nTransformer uses self-attention to make neural translation stronger.\n\nReal-time AI Translation\nReal-time AI Translation often uses NMT for speech and subtitles.", "relations": { "seq2seq": { "label": "often modeled with …", "note": "Seq2Seq encodes the source sentence, then writes the translation." }, "bahdanau-attention": { "label": "aligns meaning with …", "note": "Attention helps long sentences stay on track." }, "transformer": { "label": "gets faster and better with …", "note": "Self-attention gave translation a big jump." }, "real-time-ai-translation": { "label": "powers …", "note": "Real-time translation connects the output to speech and subtitles." } } }, "zh": { "fullName": "神经机器翻译", "factExplain": "用神经网络完成不同语言间的自动翻译。", "humanExplain": "NMT像影视配音演员：不逐字硬翻，先吃透整段台词，再用另一种语言重新讲一遍。\n\n它用于网页、字幕和跨境客服；术语和冷门语种仍要校对。", "humanExplainDisplay": "NMT像==影视配音演员==：\n不逐字硬翻，\n先吃透整段台词，\n再用另一种语言==重新讲一遍==。\n\n它用于网页、字幕和跨境客服；\n术语和冷门语种，\n仍要校对。", "relationsNarrative": "Seq2Seq\nSeq2Seq 是早期神经翻译的主力框架。\n\nBahdanau Attention\n它让模型翻长句时能对准关键原词。\n\nTransformer\n它用自注意力把神经翻译推向更强效果。\n\nReal-time AI Translation\n实时翻译常把它接到语音、字幕场景里。", "relations": { "seq2seq": { "label": "常用…建模", "note": "它把原句编码后再生成译文。" }, "bahdanau-attention": { "label": "用…对齐词义", "note": "注意力缓解长句翻译跑偏。" }, "transformer": { "label": "借…提速提质", "note": "自注意力让翻译效果大幅跃升。" }, "real-time-ai-translation": { "label": "支撑…", "note": "实时翻译把译文接到语音字幕。" } } } } }, { "id": "neural-network", "name": "Neural-network", "layer": "L1", "era": "1950s", "publishedAt": "2026-05-23T09:00:00Z", "relations": [ { "to": "parameter" }, { "to": "transformer" }, { "to": "diffusion" }, { "to": "embedding" } ], "i18n": { "en": { "fullName": "Neural Network", "factExplain": "A model made of layered nodes that learn patterns from data.", "humanExplain": "A neural network is like a school cafeteria line for guesses. Each station adds one clue. The last station yells, “Cat!”\n\nIt helps AI see pictures, use words, and make new stuff. It is the basic frame inside many AI models.", "humanExplainDisplay": "A neural network is like a ==school cafeteria line== for guesses.\nEach station adds ==one clue==.\nThe last station yells, “Cat!”\n\nIt helps AI see pictures,\nuse words,\nand make new stuff.\nIt is the basic frame inside many AI models.", "relationsNarrative": "Parameter\nParameters decide the behavior a neural network learns after training.\n\nTransformer\nA Transformer helps a neural network handle long sequences.\n\nDiffusion\nDiffusion uses a neural network to learn how to rebuild content from noise.\n\nEmbedding\nAn Embedding is a meaning pattern learned by a neural network.", "relations": { "parameter": { "label": "behavior is set by …", "note": "Parameters are learned numbers that shape how the network acts." }, "transformer": { "label": "can use … design", "note": "A Transformer is a neural network design for long sequences." }, "diffusion": { "label": "is used by …", "note": "Diffusion uses a neural network to rebuild content from noise." }, "embedding": { "label": "creates …", "note": "Embeddings are meaning patterns learned by a neural network." } } }, "zh": { "fullName": "神经网络", "factExplain": "由多层计算节点组成、能从数据中学习模式的模型结构。", "humanExplain": "神经网络像一层层打工人传纸条，每人加工一点，最后凑出答案。\n\n它是深度学习底座，支撑识图、语音和大模型。", "humanExplainDisplay": "神经网络像==一层层打工人传纸条==，\n每人加工一点，\n最后==凑出答案==。\n\n它是深度学习底座，\n支撑识图、语音和大模型。", "relationsNarrative": "Parameter\nParameter 决定 Neural-network 训练后形成的行为模式。\n\nTransformer\nTransformer 扩展了 Neural-network 处理长序列的能力。\n\nDiffusion\nDiffusion 使用 Neural-network 学习从噪声还原内容。\n\nEmbedding\nEmbedding 是 Neural-network 学到的语义表示结果。", "relations": { "parameter": { "label": "行为由…决定" }, "transformer": { "label": "包含…架构" }, "diffusion": { "label": "包含…架构" }, "embedding": { "label": "生成…" } } } } }, { "id": "neural-ode", "name": "Neural ODE", "layer": "L3", "era": "2018", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "resnet" }, { "to": "automatic-differentiation" } ], "track": "history", "i18n": { "en": { "fullName": "Neural Ordinary Differential Equation", "factExplain": "A model using a neural network to describe change over continuous time.", "humanExplain": "Neural ODE is like a volume knob, not a clicky dial. The sound glides up smoothly instead of jumping.\n\nYou meet it in data over time and physics models. The path stays smooth, but the computer works harder.", "humanExplainDisplay": "Neural ODE is like a ==volume knob==,\nnot a clicky dial.\nThe sound ==glides up smoothly==\ninstead of jumping.\n\nYou meet it in data over time\nand physics models.\nThe path stays smooth,\nbut the computer works harder.", "relationsNarrative": "Neural-network\nA Neural ODE turns layer-to-layer changes into smooth time motion.\n\nResNet\nA Neural ODE is often seen as a ResNet with endless tiny layers.\n\nAutodiff\nA Neural ODE needs Autodiff to get gradients through the solving process.", "relations": { "neural-network": { "label": "makes … continuous", "note": "It treats changes between layers as motion through continuous time." }, "resnet": { "label": "acts like the limit of …", "note": "It is often seen as a ResNet with endless tiny layers." }, "automatic-differentiation": { "label": "trains with …", "note": "Autodiff finds the gradients needed to train through solving." } } }, "zh": { "fullName": "神经常微分方程", "factExplain": "用神经网络表示连续时间微分方程的模型。", "humanExplain": "Neural ODE 像把楼梯换成滑梯：不再数台阶，每一处的坡度说了算，一路连续滑到底。\n\n适合时间序列和物理建模，过程连续但更费算。", "humanExplainDisplay": "Neural ODE 像把楼梯\n==换成滑梯==：\n不再数台阶，\n每一处的==坡度说了算==，\n一路连续滑到底。\n\n适合时间序列和物理建模，\n过程连续，\n但更费算。", "relationsNarrative": "Neural Network\n神经微分方程把网络层间变化改写为连续动态。\n\nResNet\n它常被看作 ResNet 走向无限层的连续版本。\n\nAutodiff\n训练它需要对连续求解过程计算梯度。", "relations": { "neural-network": { "label": "连续化…", "note": "把层间变化看成连续时间动态。" }, "resnet": { "label": "像…的极限", "note": "常被理解为无限层残差网络。" }, "automatic-differentiation": { "label": "借…训练", "note": "训练需对求解过程计算梯度。" } } } } }, { "id": "neural-radiance-field", "name": "NeRF", "layer": "L3", "era": "2020", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "3d-ai-generation" }, { "to": "computer-vision" }, { "to": "neural-network" } ], "track": "history", "i18n": { "en": { "fullName": "Neural Radiance Field", "factExplain": "A method that builds a 3D scene from photos taken from different angles.", "humanExplain": "Give NeRF a few phone photos of your living room. It builds a tiny 3D playset, dust bunnies included.\n\nIt helps rebuild scenes and make new camera views. The picture can look real, but training and rendering are slow.", "humanExplainDisplay": "Give NeRF a few ==phone photos==\nof your living room.\nIt builds a tiny ==3D playset==,\ndust bunnies included.\n\nIt helps rebuild scenes\nand make new camera views.\nThe picture can look real,\nbut training and rendering are slow.", "relationsNarrative": "3D AI Generation\nNeRF is a classic way to turn photos into a browsable 3D scene.\n\nComputer Vision\nNeRF uses photos from many angles to rebuild scenes and make new views.\n\nNeural-network\nNeRF stores color and density for points in space inside a neural network.", "relations": { "3d-ai-generation": { "label": "builds scenes for …", "note": "NeRF was an early path for high-detail 3D AI scenes." }, "computer-vision": { "label": "supports … reconstruction", "note": "NeRF mainly starts with photos from many angles." }, "neural-network": { "label": "stores space in …", "note": "The network holds each point's color and density." } } }, "zh": { "fullName": "神经辐射场（Neural Radiance Field）", "factExplain": "用神经网络从多视角图像重建3D场景的方法。", "humanExplain": "给NeRF几张游客照，它能在脑里搭景区沙盘：换个机位也能补拍。\n\n它用于3D重建和新视角生成；画面真，但训练渲染费劲。", "humanExplainDisplay": "给NeRF几张==游客照==，\n它能在脑里搭==景区沙盘==：\n换个机位也能补拍。\n\n它用于3D重建和新视角生成；\n画面真，\n但训练渲染费劲。", "relationsNarrative": "3D AI Generation\n神经辐射场是把照片变成可浏览3D场景的经典路线。\n\nComputer Vision\n它依赖多视角图像，解决视角合成和三维重建。\n\nNeural Network\n它用神经网络存下空间位置的颜色和密度。", "relations": { "3d-ai-generation": { "label": "生成…的场景表示", "note": "它是早期高保真3D生成路线。" }, "computer-vision": { "label": "服务…的重建任务", "note": "多视角照片是它的主要输入。" }, "neural-network": { "label": "用…表示空间", "note": "场景颜色和密度藏在网络里。" } } } } }, { "id": "neural-rendering", "name": "Neural rendering", "layer": "L2", "era": "2018", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "neural-radiance-field" }, { "to": "3d-ai-generation" }, { "to": "computer-vision" }, { "to": "diffusion" } ], "track": "today", "i18n": { "en": { "fullName": "Neural Rendering", "factExplain": "A rendering method that uses neural networks to make or rebuild images and scenes.", "humanExplain": "Neural rendering is like a video game artist with one selfie. It can paint the missing side of your head.\n\nIt can make new camera views. It can rebuild 3D scenes. Film teams use it for virtual shots.", "humanExplainDisplay": "Neural rendering is like a ==video game artist==\nwith one selfie.\nIt can paint the ==missing side==\nof your head.\n\nIt can make new camera views.\nIt can rebuild 3D scenes.\nFilm teams use it for virtual shots.", "relationsNarrative": "NeRF\nNeRF is the best-known example of neural rendering.\n\n3D AI Generation\nNeural rendering turns generated 3D results into scenes people can view.\n\nComputer Vision\nIt uses Computer Vision to understand images and space.\n\nDiffusion\nDiffusion can add texture and fill missing views.", "relations": { "neural-radiance-field": { "label": "has … as a classic example", "note": "NeRF is a classic path for neural rendering." }, "3d-ai-generation": { "label": "supports …", "note": "3D generation often needs a form the computer can render." }, "computer-vision": { "label": "depends on …", "note": "Computer Vision helps decide if the shape and texture make sense." }, "diffusion": { "label": "pairs with …", "note": "Diffusion often fills in texture and missing views." } } }, "zh": { "fullName": "神经渲染", "factExplain": "用神经网络合成或重建图像与场景的渲染方法。", "humanExplain": "神经渲染像景区导游拿一张照片讲全园：你没拍的转角，它也能补景。\n\n它用于新视角合成、3D 重建和虚拟拍摄。", "humanExplainDisplay": "神经渲染像景区导游\n拿一张照片讲全园：\n==你没拍的转角==，\n它也能补景。\n\n它用于新视角合成、\n3D 重建，\n和虚拟拍摄。", "relationsNarrative": "NeRF\nNeRF 是神经渲染最出圈的代表方法。\n\n3D AI Generation\n神经渲染常把生成结果变成可观看场景。\n\nComputer Vision\n它依赖视觉模型理解图像与空间关系。\n\nDiffusion\n扩散模型可提供纹理、视角或先验补全。", "relations": { "neural-radiance-field": { "label": "以…为代表", "note": "NeRF 是神经渲染的经典路线。" }, "3d-ai-generation": { "label": "支撑…", "note": "生成 3D 场景常需要可渲染表示。" }, "computer-vision": { "label": "依赖…", "note": "视觉理解决定几何与纹理是否靠谱。" }, "diffusion": { "label": "结合…", "note": "扩散模型常补纹理和缺失视角。" } } } } }, { "id": "neural-tangent-kernel", "name": "NTK", "layer": "L2", "era": "2018", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "kernel-method" }, { "to": "neural-network" }, { "to": "overparameterization" }, { "to": "gradient-descent" } ], "track": "history", "i18n": { "en": { "fullName": "Neural Tangent Kernel", "factExplain": "A kernel that describes how very wide neural networks change during training.", "humanExplain": "NTK is like a toy train track for a huge neural net. The engine moves, but the rails decide its path.\n\nIt helps researchers study ultra-wide networks. You meet it in theory papers on deep learning and Kernel Methods.", "humanExplainDisplay": "NTK is like a ==toy train track==\nfor a huge neural net.\nThe engine moves,\nbut the ==rails decide its path==.\n\nIt helps researchers study ultra-wide networks.\nYou meet it in theory papers\non deep learning and Kernel Methods.", "relationsNarrative": "Kernel Method\nNTK connects ultra-wide neural network training to kernel regression.\n\nNeural-network\nIt describes how a neural network trains near its starting point.\n\nOverparameterization\nThe NTK approximation works more easily when the network is very wide.\n\nGradient Descent\nIt describes how gradient descent changes the model output.", "relations": { "kernel-method": { "label": "turns networks into …", "note": "NTK approximates a wide network as kernel regression." }, "neural-network": { "label": "studies … training", "note": "It describes how wide neural networks train." }, "overparameterization": { "label": "explains behavior under …", "note": "Wider networks often make the NTK approximation more stable." }, "gradient-descent": { "label": "tracks … path", "note": "NTK tracks how gradient descent changes model outputs." } } }, "zh": { "fullName": "神经切线核", "factExplain": "描述宽神经网络训练动态的核函数。", "humanExplain": "NTK 像提前铺好的铁轨：火车再怎么使劲跑，路线在发车前就定死了。\n\n用来分析超宽网络，连接深度学习和核方法。", "humanExplainDisplay": "NTK 像提前==铺好的铁轨==：\n火车再怎么使劲跑，\n路线在==发车前就定死了==。\n\n用来分析超宽网络，\n连接深度学习，\n和核方法。", "relationsNarrative": "Kernel Method\nNTK 把超宽神经网络训练连接到核回归。\n\nNeural-network\n它刻画神经网络在初始化附近的训练动态。\n\nOverparameterization\n网络足够宽时，NTK 近似更容易成立。\n\nGradient Descent\n它描述梯度下降如何改变模型输出。", "relations": { "kernel-method": { "label": "把网络化成…", "note": "NTK 把宽网络近似成核回归。" }, "neural-network": { "label": "分析…训练", "note": "它刻画宽神经网络的训练动态。" }, "overparameterization": { "label": "解释…下的行为", "note": "网络越宽，NTK 近似通常越稳。" }, "gradient-descent": { "label": "描述…轨迹", "note": "NTK 追踪梯度下降如何改输出。" } } } } }, { "id": "neuromorphic-computing", "name": "Neuromorphic computing", "layer": "L5", "sublayer": "compute", "era": "1980s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "ai-chip" }, { "to": "edge-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Neuromorphic computing", "factExplain": "A computer design that copies how brain cells send signals.", "humanExplain": "Neuromorphic computing is like a motion-sensor porch light. It naps until a raccoon waddles by.\n\nIt helps Edge AI sensors and robots notice changes while using little power. It is not a general brain.", "humanExplainDisplay": "Neuromorphic computing is like a ==motion-sensor porch light==.\nIt naps until ==a raccoon waddles by==.\n\nIt helps Edge AI sensors and robots\nnotice changes while using little power.\nIt is not a general brain.", "relationsNarrative": "Neural-network\nNeuromorphic computing turns neuron-style signals into a computer design.\n\nAI chip\nMany neuromorphic systems use special chips for brain-like circuits.\n\nEdge AI\nIts low-power design helps edge devices react in real time.", "relations": { "neural-network": { "label": "takes ideas from …", "note": "It builds neuron-style computing into hardware." }, "ai-chip": { "label": "often runs on …", "note": "Special AI chips can hold brain-like circuits." }, "edge-ai": { "label": "helps … save power", "note": "Low power helps edge devices sense things in real time." } } }, "zh": { "fullName": "神经形态计算", "factExplain": "模仿大脑神经元工作方式的计算架构。", "humanExplain": "神经形态计算像小区声控灯：没人路过就装睡，脚步一响才亮。\n\n它适合低功耗感知、机器人和端侧设备，但不是通用大脑。", "humanExplainDisplay": "神经形态计算像\n==小区声控灯==：\n没人路过就装睡，\n脚步一响才亮。\n\n它适合低功耗感知、\n机器人和端侧设备，\n但不是通用大脑。", "relationsNarrative": "Neural-network\n它把神经元式信号处理做成计算架构。\n\nAI Chip\n许多方案依赖专用芯片承载仿脑电路。\n\nEdge AI\n低功耗、事件驱动让端侧设备更受益。", "relations": { "neural-network": { "label": "借鉴…的灵感", "note": "把神经元思路做进计算硬件。" }, "ai-chip": { "label": "常落到…上", "note": "专用芯片承载这种仿脑计算。" }, "edge-ai": { "label": "让…更省电", "note": "低功耗适合端侧实时感知。" } } } } }, { "id": "no-free-lunch-theorem", "name": "No Free Lunch Theorem", "layer": "L2", "era": "1997", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "inductive-bias" }, { "to": "statistical-learning-theory" }, { "to": "cross-validation" } ], "track": "history", "i18n": { "en": { "fullName": "No Free Lunch Theorem", "factExplain": "A theorem that no learning algorithm is best for every possible task.", "humanExplain": "No Free Lunch is one pair of shoes for every sport. Flip-flops work at the pool, not in ice hockey.\n\nIt matters when you choose models or read benchmarks. For your real task, let the validation set decide.", "humanExplainDisplay": "No Free Lunch is ==one pair of shoes==\nfor every sport.\nFlip-flops work at the pool,\nnot in ==ice hockey==.\n\nIt matters when you choose models\nor read benchmarks.\nFor your real task,\nlet the validation set decide.", "relationsNarrative": "Inductive Bias\nNo Free Lunch shows an algorithm wins by matching a problem well.\n\nSLT\nNo Free Lunch is a learning theory result about algorithm limits.\n\nCross-Validation\nNo universal winner exists, so validation results must guide model choice.", "relations": { "inductive-bias": { "label": "shows need for …", "note": "Without a built-in preference, no algorithm wins everywhere." }, "statistical-learning-theory": { "label": "is a result in …", "note": "It describes a basic limit of learning algorithms." }, "cross-validation": { "label": "pushes model choice toward …", "note": "For a real task, the validation set should decide." } } }, "zh": { "fullName": "没有免费午餐定理", "factExplain": "说明不存在对所有任务都最优的学习算法。", "humanExplain": "没有免费午餐像点外卖：川菜馆封神的套餐，拿去喂广东胃可能翻车。\n\n用于模型选型和基准测试；具体任务，还得验证集拍板。", "humanExplainDisplay": "没有免费午餐像点外卖：\n==川菜馆封神==的套餐，\n拿去喂广东胃\n可能==翻车==。\n\n用于模型选型和基准测试；\n具体任务，\n还得验证集拍板。", "relationsNarrative": "Inductive Bias\n没有免费午餐定理说明，算法优势来自问题偏好。\n\nStatistical Learning Theory\n它是学习理论中关于算法极限的结论。\n\nCross-Validation\n既然没有万能算法，选型就要看验证表现。", "relations": { "inductive-bias": { "label": "凸显…必要性", "note": "没有偏好，就没有普适优势。" }, "statistical-learning-theory": { "label": "属于…结论", "note": "它刻画学习算法的根本限制。" }, "cross-validation": { "label": "促使用…选型", "note": "具体任务上仍要验证集说话。" } } } } }, { "id": "nonmonotonic-reasoning", "name": "NMR", "layer": "L2", "era": "1980", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "belief-state" }, { "to": "knowledge-representation" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Nonmonotonic Reasoning", "factExplain": "A way to reason and take back old answers after new facts arrive.", "humanExplain": "It is like a park picnic plan. Blue sky means sandwiches, but thunder means couch pizza.\n\nIt helps with answers and plans when facts change. It also helps AI decisions.", "humanExplainDisplay": "It is like a ==park picnic plan==.\nBlue sky means sandwiches,\nbut thunder means ==couch pizza==.\n\nIt helps with answers and plans\nwhen facts change.\nIt also helps AI decisions.", "relationsNarrative": "Belief State\nNew information pushes the Belief State to update.\n\nKR\nNonmonotonic Reasoning needs KR to show rules and exceptions.\n\nAgent\nAn Agent uses this reasoning to change its mind when the world changes.", "relations": { "belief-state": { "label": "updates …", "note": "New evidence can rewrite the Belief State." }, "knowledge-representation": { "label": "uses … for rules", "note": "KR must make rules and exceptions clear." }, "agent": { "label": "helps … change its mind", "note": "An Agent may drop an old answer when the world changes." } } }, "zh": { "fullName": "Nonmonotonic Reasoning／非单调推理", "factExplain": "可在新信息出现后撤回原结论的推理方式。", "humanExplain": "它像天气预报改行程：早上还说能露营，下午一看暴雨云压过来，立马收帐篷改订民宿。\n\n适合信息会变的问答、规划、决策；重点是能及时改判。", "humanExplainDisplay": "它像天气预报改行程：\n早上还说==能露营==，\n下午一看暴雨云压过来，\n立马收帐篷，\n改订==民宿==。\n\n适合信息会变的问答、\n规划、决策；\n重点是能及时改判。", "relationsNarrative": "Belief State\n新信息进入后，它会推动信念状态随之更新。\n\nKnowledge Representation\n它常建立在能表达规则与例外的知识表示上。\n\nAgent\nAgent 遇到环境变化时，需要这种可改判的推理。", "relations": { "belief-state": { "label": "会更新…", "note": "新证据出现后，信念状态要改写。" }, "knowledge-representation": { "label": "常依赖…表达", "note": "要先把规则和例外表示清楚。" }, "agent": { "label": "帮助…改主意", "note": "环境变化时，智能体需撤回旧判断。" } } } } }, { "id": "normalizing-flow", "name": "Normalizing Flow", "layer": "L3", "era": "2014", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "variational-inference" }, { "to": "diffusion" } ], "track": "history", "i18n": { "en": { "fullName": "Normalizing Flow", "factExplain": "A generative model that turns simple randomness into complex data using reversible steps.", "humanExplain": "Normalizing Flow is like a perfect Play-Doh machine. It squishes a plain blob into a unicorn, then can unsquish it back.\n\nIt can make new samples and score exact odds. Its tricks must stay reversible, so the design has tight rules.", "humanExplainDisplay": "Normalizing Flow is like a ==perfect Play-Doh machine==.\nIt squishes a plain blob into a unicorn,\nthen can ==unsquish it back==.\n\nIt can make new samples\nand score exact odds.\nIts tricks must stay reversible,\nso the design has tight rules.", "relationsNarrative": "Generative Model\nNormalizing Flow is a Generative Model that can compute probabilities directly.\n\nVI\nNormalizing Flow can make the posterior guess in VI more flexible.\n\nDiffusion\nNormalizing Flow and Diffusion both generate samples, but they follow different paths.", "relations": { "generative-model": { "label": "is a kind of …", "note": "It is a Generative Model with exact probability scores." }, "variational-inference": { "label": "makes … more flexible", "note": "Reversible steps make the VI posterior guess more flexible." }, "diffusion": { "label": "takes a different path from …", "note": "Both generate samples, but they train in different ways." } } }, "zh": { "fullName": "归一化流", "factExplain": "通过可逆变换建模复杂分布的生成模型。", "humanExplain": "归一化流像拉糖人：糖坯拧成龙凤，还能按手法原路捋回去。\n\n用于生成和密度估计，概率好算，但结构受约束。", "humanExplainDisplay": "归一化流像拉糖人：\n糖坯拧成==龙凤==，\n还能按手法\n==原路捋回去==。\n\n用于生成和密度估计，\n概率好算，\n但结构受约束。", "relationsNarrative": "Generative Model\n归一化流是一类可显式计算概率的生成模型。\n\nVariational Inference\n它常用来增强变分推断里的后验近似。\n\nDiffusion\n二者都能生成样本，但建模路径不同。", "relations": { "generative-model": { "label": "属于…", "note": "它是可显式算概率的生成模型。" }, "variational-inference": { "label": "增强…后验", "note": "可逆变换让近似后验更灵活。" }, "diffusion": { "label": "对比…路线", "note": "两者都生成样本，训练取舍不同。" } } } } }, { "id": "nouvelle-ai", "name": "Nouvelle AI", "layer": "L1", "era": "1986", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "robotics" }, { "to": "symbolic-ai" } ], "track": "history", "seo": { "zh": { "title": "新人工智能 是什么?先蹬再拐,一文看懂 — AI Rookies", "description": "一种强调具身行动、反对纯符号推理的 AI 范式。人话版一分钟讲明白。" }, "en": { "title": "What Is Nouvelle AI? Wobble Before You Steer", "description": "An AI approach that favors bodies and action over pure symbol rules. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Nouvelle AI", "factExplain": "An AI approach that favors bodies and action over pure symbol rules.", "humanExplain": "Nouvelle AI is a kid learning a skateboard. No physics lecture first. Just wobble, bump a trash can, then steer better.\n\nIt says intelligence starts by doing things in the world. You meet this idea in robots and Embodied AI.", "humanExplainDisplay": "Nouvelle AI is a kid learning a ==skateboard==.\nNo physics lecture first.\nJust ==wobble, bump a trash can==,\nthen steer better.\n\nIt says intelligence starts\nby doing things in the world.\nYou meet this idea\nin robots and Embodied AI.", "relationsNarrative": "Subsumption Architecture\nNouvelle AI used layered behavior control to make robots act.\n\nEmbodied AI\nNouvelle AI inspired Embodied AI to link thinking with body and world.\n\nRobotics\nNouvelle AI first took root in mobile robots, not paper reasoning.\n\nSymbolic AI\nNouvelle AI pushed back on explaining intelligence with symbols alone.", "relations": { "embodied-ai": { "label": "inspired …", "note": "Embodied AI keeps its idea of thinking through action." }, "robotics": { "label": "took root in …", "note": "Mobile robots were its first real test bed." }, "symbolic-ai": { "label": "pushed back on …", "note": "It moved intelligence away from pure symbol rules and back to the real world." } } }, "zh": { "fullName": "新人工智能", "factExplain": "一种强调具身行动、反对纯符号推理的 AI 范式。", "humanExplain": "新派AI像孩子学骑车：不先背力学大部头，先蹬起来，摔两下就会拐。\n\n影响机器人和具身智能，主张先行动再谈推理。", "humanExplainDisplay": "新派AI像孩子学骑车：\n不先背==力学大部头==，\n先==蹬起来==，\n摔两下就会拐。\n\n影响机器人和具身智能，\n主张先行动再谈推理。", "relationsNarrative": "Subsumption Architecture\n它用分层行为控制，把新派 AI 落到机器人上。\n\nEmbodied AI\n它启发具身智能：智能来自身体与环境互动。\n\nRobotics\n它最早扎根移动机器人，而非纸面推理。\n\nSymbolic AI\n它反对只靠符号建模解释智能。", "relations": { "embodied-ai": { "label": "启发…", "note": "具身智能延续了边行动边认知。" }, "robotics": { "label": "扎根于…", "note": "移动机器人是它最早的实验场。" }, "symbolic-ai": { "label": "反对只靠…", "note": "它把智能从符号推理拉回现实。" } } } } }, { "id": "npu-neural-processing-unit", "name": "NPU", "layer": "L5", "sublayer": "compute", "era": "2010s", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "ai-chip" }, { "to": "gpu" }, { "to": "edge-ai" }, { "to": "tpu" } ], "track": "today", "i18n": { "en": { "fullName": "Neural Processing Unit", "factExplain": "A chip built to speed up neural network math.", "humanExplain": "An NPU is like a pizza oven. It does one job fast, without heating the whole kitchen.\n\nIt speeds up neural network math inside phones, laptops, and cars. Face unlock, translation, and photo cleanup can use less battery.", "humanExplainDisplay": "An NPU is like a ==pizza oven==.\nIt does ==one job fast==,\nwithout heating the whole kitchen.\n\nIt speeds up neural network math\ninside phones, laptops, and cars.\nFace unlock, translation, and photo cleanup\ncan use less battery.", "relationsNarrative": "AI chip\nAn NPU is a type of AI chip for neural network math.\n\nGPU\nA GPU is more general. An NPU is more focused and uses less power for AI model runs.\n\nEdge AI\nNPUs help Edge AI run faster and use less power.\n\nTPU\nA TPU and an NPU are both special chips for tensor math.", "relations": { "ai-chip": { "label": "is a kind of …", "note": "An NPU is an AI chip built for neural networks." }, "gpu": { "label": "differs from …", "note": "An NPU is more focused than a GPU and uses less power for AI model runs." }, "edge-ai": { "label": "speeds up …", "note": "NPUs help Edge AI run faster without sending as much to the cloud." }, "tpu": { "label": "is specialized like …", "note": "TPUs and NPUs both speed up tensor math." } } }, "zh": { "fullName": "Neural Processing Unit，神经网络处理器", "factExplain": "专为神经网络计算设计的加速芯片。", "humanExplain": "NPU 像高考数学专练生：不抢语文英语，只刷 AI 计算题，快还省草稿纸。\n\n装进手机、电脑和车机，本地识别、翻译、拍照增强更省电。", "humanExplainDisplay": "NPU 像==高考数学专练生==：\n不抢语文英语，\n只刷 AI 计算题，\n==快还省草稿纸==。\n\n装进手机、电脑和车机，\n本地识别、翻译、拍照增强更省电。", "relationsNarrative": "AI Chip\nNPU 是 AI 芯片的一类，专攻神经网络计算。\n\nGPU\nGPU 更通用，NPU 更偏低功耗专用推理。\n\nEdge AI\nNPU 让端侧 AI 能更快、更省电地运行。\n\nTPU\nTPU 和 NPU 都是为张量计算加速的专用芯片。", "relations": { "ai-chip": { "label": "属于…", "note": "NPU 是面向神经网络的 AI 芯片。" }, "gpu": { "label": "对比…", "note": "NPU 更偏低功耗专用推理。" }, "edge-ai": { "label": "加速…", "note": "NPU 让端侧 AI 少依赖云端。" }, "tpu": { "label": "同属专用…", "note": "二者都为张量计算加速。" } } } } }, { "id": "object-detection", "name": "Object Detection", "layer": "L4", "era": "1990s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "image-classification" }, { "to": "cnn" }, { "to": "coco-dataset" } ], "track": "history", "i18n": { "en": { "fullName": "Object Detection", "factExplain": "A vision task that names objects and finds them in an image.", "humanExplain": "Object detection is I Spy with a highlighter. It says “dog” and draws a box around the muddy little suspect.\n\nIt finds objects and their spots in pictures. Cars use it. Security cameras use it. Photo apps use it.", "humanExplainDisplay": "Object detection is ==I Spy== with a highlighter.\nIt says “dog”\nand draws a ==box== around the muddy little suspect.\n\nIt finds objects and their spots in pictures.\nCars use it.\nSecurity cameras use it.\nPhoto apps use it.", "relationsNarrative": "Computer Vision\nObject detection is a core task in Computer Vision.\n\nImage Class.\nIt names what an image shows, while detection also finds where the object is.\n\nCNN\nMany classic detectors use CNNs to read useful image features.\n\nCOCO Dataset\nCOCO Dataset is often used to train and test object detection models.", "relations": { "computer-vision": { "label": "is a core … task", "note": "Object detection is a core task in Computer Vision." }, "image-classification": { "label": "goes beyond …", "note": "Image Class. names the object, but detection also finds its location." }, "cnn": { "label": "often uses …", "note": "Many classic detectors used CNNs to pull features from images." }, "coco-dataset": { "label": "is tested on …", "note": "COCO Dataset is often used to train and test detection models." } } }, "zh": { "fullName": "目标检测", "factExplain": "识别图像中物体类别和位置的视觉任务。", "humanExplain": "目标检测像相机的人脸对焦框：不光看见画面里有人，还把每张脸的位置框出来。\n\n它用于自动驾驶、安防和相册，负责找物体与位置。", "humanExplainDisplay": "目标检测像相机对焦框：\n==不光看见有人==，\n还把每张脸\n在画面里==框出来==。\n\n它用于自动驾驶、安防和相册，\n负责找物体\n与位置。", "relationsNarrative": "Computer Vision\n目标检测是计算机视觉里的核心任务。\n\nImage Classification\n图像分类只认是什么，目标检测还要定位。\n\nCNN\n许多经典检测器用 CNN 提取图像特征。\n\nCOCO Dataset\nCOCO Dataset 常用于训练和评测检测模型。", "relations": { "computer-vision": { "label": "属于…任务", "note": "目标检测是计算机视觉的核心任务。" }, "image-classification": { "label": "扩展…", "note": "它不只认类别，还要找出位置。" }, "cnn": { "label": "常用…提特征", "note": "早期主流检测器大量依赖 CNN。" }, "coco-dataset": { "label": "用…评测", "note": "COCO 常用于训练和评测检测模型。" } } } } }, { "id": "ocr", "name": "OCR", "layer": "L5", "sublayer": "architecture", "era": "1970s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "document-parsing" }, { "to": "multimodal" }, { "to": "ai-file-upload" } ], "track": "today", "seo": { "en": { "title": "What Is OCR? How AI Reads Text From Images", "description": "OCR turns words in photos and scans into text a computer can process — the starting point of document automation. Explained simply, with a concrete analogy." }, "zh": { "title": "OCR 是什么?让电脑认出图片里的字,一文看懂 — AI Rookies", "description": "票据录入、合同整理、截图取字都靠它:把图上的字认成可处理的文本。文档自动化的起点,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Optical Character Recognition", "factExplain": "A technology that turns words in images or scans into text a computer can use.", "humanExplain": "OCR is a phone camera with reading glasses. It squints at a blurry receipt and turns the printed mess into real text.\n\nYou meet it in receipt apps and scanned contracts. It turns screenshots into copyable words too.", "humanExplainDisplay": "OCR is a phone camera with ==reading glasses==.\nIt squints at a blurry receipt\nand turns the printed mess into ==real text==.\n\nYou meet it in receipt apps and scanned contracts.\nIt turns screenshots into copyable words too.", "relationsNarrative": "Computer Vision\nOCR is one of the classic real-world tasks in Computer Vision.\n\nDocument parsing\nDocument parsing often uses OCR first, then pulls out fields and structure.\n\nMultimodal AI\nOCR turns words in images into text for Multimodal AI to read.\n\nAI File Upload\nAfter a user uploads an image or PDF, OCR often extracts the text first.", "relations": { "computer-vision": { "label": "is a … task", "note": "OCR is a classic Computer Vision job." }, "document-parsing": { "label": "often comes before …", "note": "OCR reads the words before fields and structure get pulled out." }, "multimodal": { "label": "feeds text to …", "note": "OCR gives Multimodal AI readable text from a picture." }, "ai-file-upload": { "label": "reads content in …", "note": "After a scan is uploaded, OCR often pulls out the text." } } }, "zh": { "fullName": "光学字符识别", "factExplain": "把图片或扫描件里的文字识别成可处理文本的技术。", "humanExplain": "它像考试后借学霸笔记誊答案的人：先盯着那页歪歪扭扭的字，硬是把==图上的字认出来==，再交给电脑==接着处理==。\n\n常用于票据录入、合同整理和截图取字，是文档自动化起点。", "humanExplainDisplay": "它像考试后借学霸笔记誊答案的人：\n先盯着那页歪歪扭扭的字，\n硬是把==图上的字认出来==，\n再交给电脑==接着处理==。\n\n常用于票据录入、\n合同整理和截图取字，\n是文档自动化起点。", "relationsNarrative": "Computer Vision\n它是计算机视觉里最经典的落地任务之一。\n\nDocument parsing\n文档解析常先靠它识字，再做结构化提取。\n\nMultimodal\n它把图里的字变成文本，方便多模态模型理解。\n\nAi-file-upload\n用户上传图片或 PDF 后，常先用它提取文字。", "relations": { "computer-vision": { "label": "属于…应用", "note": "它是计算机视觉的经典任务之一。" }, "document-parsing": { "label": "常作…前置", "note": "先识字，再抽字段和结构。" }, "multimodal": { "label": "给…提供文字", "note": "把图里文字变成可读输入。" }, "ai-file-upload": { "label": "处理…内容", "note": "上传扫描件后常靠它识字。" } } } } }, { "id": "off-policy-learning", "name": "Off-policy-learning", "layer": "L2", "era": "1989", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "q-learning" }, { "to": "reinforcement-learning" }, { "to": "deep-q-network" }, { "to": "temporal-difference-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Off-Policy Learning", "factExplain": "A way for AI to learn from data made by another policy.", "humanExplain": "Off-policy learning is like studying Mario Kart from your cousin’s messy replays. You did not drive those laps, but you learn when to brake.\n\nThe AI learns from old runs or other players’ moves. It shows up in recommenders, games, and robots.", "humanExplainDisplay": "Off-policy learning is like studying Mario Kart\nfrom your cousin’s ==messy replays==.\nYou did not drive those laps,\nbut you learn ==when to brake==.\n\nThe AI learns from old runs\nor other players’ moves.\nIt shows up in recommenders,\ngames, and robots.", "relationsNarrative": "Q-Learning\nQ-Learning is the classic example of off-policy learning.\n\nRL\nOff-policy learning is part of RL and can use old experience.\n\nDeep Q-Network\nDeep Q-Network uses replay data to learn from old runs again.\n\nTD Learning\nOff-policy learning often uses TD Learning to update value guesses.", "relations": { "q-learning": { "label": "often appears as …", "note": "Q-Learning is a classic off-policy method." }, "reinforcement-learning": { "label": "belongs to …", "note": "Off-policy learning is a core training path in RL." }, "deep-q-network": { "label": "trains …", "note": "DQN learns from replayed old data." }, "temporal-difference-learning": { "label": "updates with …", "note": "TD Learning can estimate value from past experience." } } }, "zh": { "fullName": "Off-Policy Learning／离策略学习", "factExplain": "用非当前策略生成的数据来学习策略的方法。", "humanExplain": "离策略像刷网购攻略，不用自己把每家坑都踩完，看别人的买家秀和避雷贴，也能少交学费。\n\n常用于复用旧数据训练，在推荐、游戏和机器人里更省样本。", "humanExplainDisplay": "离策略像刷网购攻略，\n不用自己把每家坑\n都踩完，\n看别人的==买家秀==\n和避雷贴，\n也能少交\n==学费==。\n\n常用于复用旧数据训练，\n在推荐、游戏\n和机器人里更省样本。", "relationsNarrative": "Q-Learning\nQ-Learning 是离策略学习里最经典的代表方法。\n\nReinforcement Learning\n它属于强化学习，用历史经验而非只靠当前尝试。\n\nDeep Q-Network\nDQN 依赖经验回放，让模型用旧数据反复学习。\n\nTD Learning\n它常配合时序差分更新价值估计。", "relations": { "q-learning": { "label": "常见于…", "note": "Q-Learning 是典型离策略方法。" }, "reinforcement-learning": { "label": "属于…范式", "note": "它是强化学习里的核心训练路线。" }, "deep-q-network": { "label": "支撑…训练", "note": "DQN 用回放数据离线更新策略。" }, "temporal-difference-learning": { "label": "常结合…更新", "note": "常用时序差分从旧经验估计价值。" } } } } }, { "id": "ollama", "name": "Ollama", "layer": "L5", "sublayer": "product", "era": "2023", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "local-llm" }, { "to": "gguf" }, { "to": "on-premise-ai" }, { "to": "llm" } ], "track": "history", "i18n": { "en": { "fullName": "Ollama", "factExplain": "A tool for running large language models on your own device.", "humanExplain": "Ollama is like a home generator for AI models. Your laptop makes its own power, no grid needed.\n\nPeople use it to run models locally and test apps. It also helps private data stay put.", "humanExplainDisplay": "Ollama is like a ==home generator==\nfor AI models.\nYour laptop ==makes its own power==,\nno grid needed.\n\nPeople use it to run models locally\nand test apps.\nIt also helps private data stay put.", "relationsNarrative": "Local-LLM\nOllama is one common way to run a Local-LLM.\n\nGGUF\nMany local models use GGUF files with Ollama.\n\nOn-premise AI\nOllama is often a light first step for On-premise AI.\n\nLLM\nOllama brings LLMs from the cloud to your own device.", "relations": { "local-llm": { "label": "often runs …", "note": "It makes Local-LLMs easier to install and use." }, "gguf": { "label": "loads models with …", "note": "Many local models use GGUF files." }, "on-premise-ai": { "label": "starts light …", "note": "It makes private deployment easier to start." }, "llm": { "label": "brings … local", "note": "It lets LLMs run outside cloud services." } } }, "zh": { "fullName": "本地大模型运行工具", "factExplain": "一个让大模型在本地设备上运行的工具。", "humanExplain": "它像家里阳台上的一台小发电机：不用总连市电网，想用电自己就发，断了网也照样转。\n\n常用来本地跑模型、开发测试，也适合隐私敏感场景。", "humanExplainDisplay": "它像家里阳台上的\n一台==小发电机==：\n不用总连市电网，\n想用电==自己就发==，\n断了网也照样转。\n\n常用来本地跑模型、\n开发测试，\n也适合隐私敏感场景。", "relationsNarrative": "Local-LLM\n它是本地大模型最常见的运行入口之一。\n\nGGUF\n很多本地模型会用 GGUF 格式给它加载。\n\nOn-premise AI\n它常被拿来做私有部署的轻量起步方案。\n\nLLM\n它把原本常在云端跑的大模型搬到个人设备上。", "relations": { "local-llm": { "label": "常用来跑…", "note": "它让本地大模型更容易装和用。" }, "gguf": { "label": "常搭配…加载", "note": "很多本地模型会用这种文件格式。" }, "on-premise-ai": { "label": "是…的轻量入口", "note": "它让私有部署更容易起步。" }, "llm": { "label": "把…搬到本地", "note": "让大模型不只待在云端服务里。" } } } } }, { "id": "on-policy-learning", "name": "On-Policy Learning", "layer": "L2", "era": "1988", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "off-policy-learning" }, { "to": "policy-gradient" }, { "to": "proximal-policy-optimization" } ], "track": "history", "i18n": { "en": { "fullName": "On-Policy Learning", "factExplain": "A reinforcement learning method using its current policy’s own trial data.", "humanExplain": "On-policy learning is like a basketball player watching their own game tape. They fix their own air ball, not the kid on the next court.\n\nThe AI tests its current plan, then updates that same plan. You meet it in Policy Gradient and PPO, and it is steady but needs many practice runs.", "humanExplainDisplay": "On-policy learning is like a basketball player\nwatching their ==own game tape==.\nThey fix their ==own air ball==,\nnot the kid on the next court.\n\nThe AI tests its current plan,\nthen updates that same plan.\nYou meet it in Policy Gradient and PPO,\nand it is steady but needs many practice runs.", "relationsNarrative": "RL\nOn-policy learning is one way to update a policy in RL.\n\nOff-policy-learning\nThe main difference is whether the data came from the current policy.\n\nPolicy Gradient\nMany Policy Gradient methods use samples from the current policy.\n\nPPO\nPPO is a classic on-policy training method.", "relations": { "reinforcement-learning": { "label": "is a type of …", "note": "It is one way to update a policy in RL." }, "off-policy-learning": { "label": "contrasts with …", "note": "The key difference is whose policy made the data." }, "policy-gradient": { "label": "is often used in …", "note": "Many Policy Gradient methods sample with the current policy." }, "proximal-policy-optimization": { "label": "supports … training", "note": "PPO is a classic on-policy training method." } } }, "zh": { "fullName": "同策略学习", "factExplain": "用当前策略采样并更新同一策略的强化学习方法。", "humanExplain": "同策略学习像棋手复盘自己这盘棋：哪步臭就改哪步，不照搬旁桌。\n\n用于策略梯度、PPO，更新稳，但更费样本。", "humanExplainDisplay": "同策略学习像棋手\n复盘自己这盘棋：\n==哪步臭就改哪步==，\n不照搬旁桌。\n\n用于策略梯度、PPO，\n更新稳，\n但更费样本。", "relationsNarrative": "RL\n同策略学习是强化学习中更新策略的一类方法。\n\nOff-policy Learning\n两者核心区别，是数据是否来自当前策略。\n\nPolicy Gradient\n许多策略梯度方法依赖当前策略采样。\n\nPPO\nPPO 是典型的同策略训练算法。", "relations": { "reinforcement-learning": { "label": "属于…方法", "note": "它是强化学习里更新策略的一类做法。" }, "off-policy-learning": { "label": "对比…", "note": "两者区别在数据来自谁的策略。" }, "policy-gradient": { "label": "常用于…", "note": "许多策略梯度方法按当前策略采样。" }, "proximal-policy-optimization": { "label": "支撑…训练", "note": "PPO 是典型的同策略训练算法。" } } } } }, { "id": "on-premise-ai", "name": "On-premise AI", "layer": "L5", "sublayer": "product", "era": "2024", "publishedAt": "2026-05-29T16:08:01.212Z", "relations": [ { "to": "local-llm" }, { "to": "data-privacy" }, { "to": "ai-regulation" }, { "to": "gpu" } ], "track": "today", "i18n": { "en": { "fullName": "On-premise AI", "factExplain": "AI systems run on a company’s own servers or private network.", "humanExplain": "On-premise AI is like keeping the family cookie jar at home. No cloud landlord gets a spare key.\n\nBanks and hospitals use it, so data stays inside. Government offices use it to meet strict rules.", "humanExplainDisplay": "On-premise AI is like keeping the\n==family cookie jar at home==.\nNo cloud landlord gets a\n==spare key==.\n\nBanks and hospitals use it,\nso data stays inside.\nGovernment offices use it\nto meet strict rules.", "relationsNarrative": "Local-LLM\nOn-premise AI often uses a Local-LLM inside the company’s own network.\n\nData-privacy\nOn-premise AI keeps sensitive data inside the private network when possible.\n\nAI-regulation\nOn-premise AI helps regulated teams meet stricter rules.\n\nGPU\nOn-premise AI needs local GPU power to run the model well.", "relations": { "local-llm": { "label": "often runs with …", "note": "A Local-LLM is a common choice for private deployment." }, "data-privacy": { "label": "helps protect …", "note": "Data stays inside the private network, so privacy is easier to manage." }, "ai-regulation": { "label": "helps meet …", "note": "On-premise AI can help with strict industry rules." }, "gpu": { "label": "needs … for power", "note": "Local AI often needs its own GPUs to train or run models." } } }, "zh": { "fullName": "本地部署 AI", "factExplain": "把 AI 系统部署在自有机房或内网环境中运行。", "humanExplain": "本地部署像把 AI 请进公司机房：不住云端酒店，门禁卡自己管。\n\n它适合敏感数据、合规要求高的企业，但要自己养机器和运维。", "humanExplainDisplay": "本地部署像==把 AI 请进公司机房==：\n==不住云端酒店==，\n门禁卡自己管。\n\n它适合敏感数据、合规要求高的企业，\n但要自己养机器和运维。", "relationsNarrative": "Local-LLM\nOn-premise AI 常搭配 Local-LLM 落地，把模型直接部署在企业自有环境中。\n\nData-privacy\nOn-premise AI 的核心吸引力之一，就是让敏感数据尽量留在内网里处理。\n\nAI-regulation\n在金融、政务、医疗等场景里，On-premise AI 常被用来满足更严格的合规要求。\n\nGPU\nOn-premise AI 不是把模型搬回来就完事，还得有 GPU 等本地算力撑住运行。", "relations": { "local-llm": { "label": "常搭配…落地", "note": "本地模型是私有部署的常见选择。" }, "data-privacy": { "label": "用于保障…", "note": "数据不出内网，更容易满足隐私要求。" }, "ai-regulation": { "label": "帮助满足…", "note": "私有部署常用于应对行业合规要求。" }, "gpu": { "label": "依赖…支撑算力", "note": "本地部署通常要自备训练或推理硬件。" } } } } }, { "id": "one-person-company", "name": "One-person Company", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "ai-native-organization" }, { "to": "ai-productivity" }, { "to": "ai-app-builder" } ], "track": "today", "i18n": { "en": { "fullName": "AI One-person Company", "factExplain": "A company where one person uses AI to handle most business jobs.", "humanExplain": "A one-person company is like a food truck with one human at the window. The AI crew preps the tacos and counts the cash.\n\nYou see it in solo content work, online stores, or tiny software startups. It boosts your output, but it does not take the risk for you.", "humanExplainDisplay": "A one-person company is like\na ==food truck==\nwith one human at the window.\nThe ==AI crew== preps the tacos\nand counts the cash.\n\nYou see it in solo content work,\nonline stores,\nor tiny software startups.\nIt boosts your output,\nbut it does not take the risk for you.", "relationsNarrative": "Agent\nAn Agent lets one person hand off some work to virtual staff.\n\nAI-native organization\nA one-person company is an AI-native organization shrunk to one person.\n\nAI Productivity\nAI Productivity lets one person cover more company jobs.\n\nAI App Builder\nAI App Builders lower the bar for building products alone.", "relations": { "agent": { "label": "uses … as virtual staff", "note": "Agents can handle some daily work and action steps." }, "ai-native-organization": { "label": "shrinks … to one person", "note": "A one-person company is the smallest AI-native organization." }, "ai-productivity": { "label": "boosts output with …", "note": "AI Productivity lets one person cover more company jobs." }, "ai-app-builder": { "label": "builds fast with …", "note": "AI App Builders make solo product work easier." } } }, "zh": { "fullName": "AI 一人公司", "factExplain": "由一个人借助 AI 完成多数公司职能的组织形态。", "humanExplain": "一人公司像煎饼摊老板带一队 AI 小工：你翻面收钱，它们和面、吆喝、记账。\n\n适合内容、电商、软件创业，放大产出，但不替你兜风险。", "humanExplainDisplay": "一人公司像==煎饼摊老板==\n带一队==AI 小工==：\n你翻面收钱，\n它们和面、吆喝、记账。\n\n适合内容、电商、软件创业，\n放大产出，\n但不替你兜风险。", "relationsNarrative": "Agent\nAgent 让一个人能把部分执行任务交给虚拟员工。\n\nAI-native Organization\n一人公司是 AI 原生组织被压缩到个人后的形态。\n\nAI Productivity\nAI 生产力提升，让单人能覆盖更多公司职能。\n\nAI App Builder\nAI 应用搭建工具降低了单人做产品的门槛。", "relations": { "agent": { "label": "用…当虚拟员工", "note": "Agent 能代做部分运营和执行。" }, "ai-native-organization": { "label": "缩小…到个人", "note": "一人公司是 AI 原生组织的极小形态。" }, "ai-productivity": { "label": "放大个人产出", "note": "生产力提升让单人覆盖更多职能。" }, "ai-app-builder": { "label": "借…快速做产品", "note": "应用搭建工具降低单人创业门槛。" } } } } }, { "id": "online-learning", "name": "Online Learning", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "reinforcement-learning" }, { "to": "federated-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Online Learning", "factExplain": "A way for a model to keep learning as new data arrives.", "humanExplain": "Online Learning is like a barista with a scary good memory. Correct your order today, and tomorrow your oat milk is safe.\n\nIt fits fast-moving feeds, ads, and fraud checks. It learns fresh patterns fast, but it can learn fresh junk too.", "humanExplainDisplay": "Online Learning is like a ==barista with a scary good memory==.\nCorrect your order today,\nand tomorrow your ==oat milk is safe==.\n\nIt fits fast-moving feeds, ads,\nand fraud checks.\nIt learns fresh patterns fast,\nbut it can learn fresh junk too.", "relationsNarrative": "Supervised Learning\nOnline Learning is an incremental update style of Supervised Learning.\n\nRL\nRL often updates online while it interacts with the world.\n\nFL\nOnline Learning can work with FL, so devices learn locally and share updates.", "relations": { "supervised-learning": { "label": "is a version of …", "note": "Both learn from examples, but Online Learning updates as data arrives." }, "reinforcement-learning": { "label": "often appears in …", "note": "Many RL systems keep updating while they interact." }, "federated-learning": { "label": "can update with …", "note": "A device can keep learning locally, then share model updates." } } }, "zh": { "fullName": "在线学习", "factExplain": "模型随着新数据到来持续更新的学习方式。", "humanExplain": "打工人最懂这套：今天刚被客户教育一顿，明天同类需求一来，手上话术就自动升级。\n\n它适合推荐、广告和风控等变化快的场景，但也容易吃进新噪声。", "humanExplainDisplay": "打工人最懂这套：\n今天刚被客户==教育一顿==，\n明天同类需求一来，\n手上话术就==自动升级==。\n\n它适合推荐、广告\n和风控等变化快的场景，\n但也容易吃进新噪声。", "relationsNarrative": "Supervised Learning\n它是监督学习的一种增量更新方式。\n\nReinforcement Learning\n强化学习常在与环境交互时在线更新。\n\nFederated Learning\n它可与联邦学习结合做分布式持续更新。", "relations": { "supervised-learning": { "label": "是…的变体", "note": "都靠样本学习，只是更新方式不同。" }, "reinforcement-learning": { "label": "常用于…场景", "note": "很多强化学习都在交互中持续更新。" }, "federated-learning": { "label": "可结合…更新", "note": "本地持续学完，再分布式汇总参数。" } } } } }, { "id": "ontology", "name": "Ontology", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "knowledge-graph" }, { "to": "rag" }, { "to": "description-logic" } ], "track": "history", "i18n": { "en": { "fullName": "Ontology", "factExplain": "A shared model for concepts, relationships, and rules in a knowledge system.", "humanExplain": "Ontology is like a bossy grocery-store map. Milk is dairy. Socks are not fruit.\n\nYou meet it in knowledge bases, search, and Q&A. It lines up data, so rules can make sense.", "humanExplainDisplay": "Ontology is like a ==bossy grocery-store map==.\n==Milk is dairy==.\nSocks are not fruit.\n\nYou meet it in knowledge bases,\nsearch, and Q&A.\nIt lines up data,\nso rules can make sense.", "relationsNarrative": "KR\nOntology is a KR method for concepts and relationships.\n\nKnowledge Graph\nAn ontology keeps a Knowledge Graph's nodes, links, and meanings consistent.\n\nRAG\nOntology helps RAG find the right sources more often.\n\nDL\nDL is often used to write ontology rules in a formal way.", "relations": { "knowledge-representation": { "label": "is core to …", "note": "Ontology writes concepts and relationships in a clear form." }, "knowledge-graph": { "label": "sets meaning for …", "note": "It gives a graph shared rules for nodes and links." }, "rag": { "label": "can sharpen … search", "note": "Shared concepts help RAG find the right material." }, "description-logic": { "label": "is often written with …", "note": "DL is a common way to state ontology rules." } } }, "zh": { "fullName": "本体论／本体", "factExplain": "一种为概念、关系和规则统一建模的知识规范。", "humanExplain": "Ontology 像商场导购图先定规矩：哪层卖什么、店和店啥关系，免得问三个人说法都不一样。\n\n常用于知识库、搜索、问答；让数据对齐，也便于按规则理解。", "humanExplainDisplay": "Ontology 像商场导购图\n先定规矩：哪层卖什么、\n店和店==啥关系==，\n免得问三个人，\n说法都==不一样==。\n\n常用于知识库、搜索、问答；\n让数据对齐，\n也便于按规则理解。", "relationsNarrative": "Knowledge Representation\n本体是知识表示里专门管概念与关系的做法。\n\nKnowledge Graph\n知识图谱常用本体统一节点、关系和含义。\n\nRAG\n本体能帮检索和问答少点答非所问。\n\nDescription Logic\n描述逻辑常被用来形式化表达本体规则。", "relations": { "knowledge-representation": { "label": "属于…核心方法", "note": "它把概念和关系明确写出来。" }, "knowledge-graph": { "label": "为…定语义", "note": "本体常给图谱提供统一 schema。" }, "rag": { "label": "可增强…检索", "note": "统一概念后更容易找准资料。" }, "description-logic": { "label": "常用…来表达", "note": "很多本体语言建立在其上。" } } } } }, { "id": "open-source-model", "name": "Open-source-model", "layer": "L3", "era": "2023", "publishedAt": "2026-05-23T10:30:00Z", "relations": [ { "to": "local-llm" }, { "to": "foundation-model" }, { "to": "fine-tuning" } ], "i18n": { "en": { "fullName": "Open-source Model", "factExplain": "An AI model whose code, weights, or use rules are open to the public.", "humanExplain": "An open-source model is like a Lego car with the instructions in the box. You can swap the wheels and spot the missing bumper.\n\nYou can run it on your own machine and change it for your job. It can save money, but updates and safety are now your problem.", "humanExplainDisplay": "An open-source model is like a ==Lego car==\nwith the ==instructions in the box==.\nYou can swap the wheels\nand spot the missing bumper.\n\nYou can run it on your own machine\nand change it for your job.\nIt can save money,\nbut updates and safety are now your problem.", "relationsNarrative": "Local-LLM\nOpen-source models give Local-LLM a model it can run on your own machine.\n\nFoundation-model\nAn open-source model often starts as an open Foundation-model.\n\nFine-tuning\nFine-tuning makes an open-source model better at one job.", "relations": { "local-llm": { "label": "helps deploy …", "note": "Open-source models give Local-LLM a model it can run on your own machine." }, "foundation-model": { "label": "often comes from …", "note": "Many open-source models are open versions of Foundation-models." }, "fine-tuning": { "label": "can be shaped by …", "note": "Fine-tuning teaches an open-source model a specific job." } } }, "zh": { "fullName": "开源模型", "factExplain": "权重、代码或使用方式开放给社区的 AI 模型。", "humanExplain": "开源模型像小区食堂把菜谱贴墙上，谁都能照着做，还能按自家口味加辣。\n\n它让个人和企业能本地部署，也方便二次开发，但会带来合规和滥用风险。", "humanExplainDisplay": "开源模型像==小区食堂把菜谱贴墙上==，\n谁都能照着做，\n还能按==自家口味加辣==。\n\n它让个人和企业能本地部署，\n也方便二次开发，\n但会带来合规和滥用风险。", "relationsNarrative": "Local-LLM\nOpen-source-model 让 Local-LLM 获得可部署的模型基础。\n\nFoundation-model\nOpen-source-model 往往来自 Foundation-model 路线的开放版本。\n\nFine-tuning\nFine-tuning 让 Open-source-model 更适合特定任务。", "relations": { "local-llm": { "label": "便于部署…" }, "foundation-model": { "label": "基于…" }, "fine-tuning": { "label": "可被…" } } } } }, { "id": "open-source-washing", "name": "Open-source Washing", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "open-weights" }, { "to": "ai-model-licenses" }, { "to": "closed-model" } ], "track": "today", "i18n": { "en": { "fullName": "Open-source Washing", "factExplain": "Marketing that makes partly open AI look fully open-source.", "humanExplain": "Open-source washing is a “free bike” sign at a yard sale. You show up and get one wheel.\n\nUse it when you read model launches or compare AI vendors. It checks if the open promise is real.", "humanExplainDisplay": "Open-source washing is a ==“free bike” sign==\nat a yard sale.\nYou show up\nand get ==one wheel==.\n\nUse it when you read model launches\nor compare AI vendors.\nIt checks if the open promise is real.", "relationsNarrative": "Open-source-model\nOpen-source washing pretends to be true open source and blurs the line.\n\nOpen weights\nOpen weights are often sold as full open source.\n\nModel-licenses\nModel licenses say if you can use, change, or share the model.\n\nClosed-source Model\nA closed-source model can use open talk to feel less locked down.", "relations": { "open-source-model": { "label": "pretends to be …", "note": "Real open-source models open the key parts." }, "open-weights": { "label": "gets confused with …", "note": "Open weights alone do not make a model fully open-source." }, "ai-model-licenses": { "label": "dodges …", "note": "A license says what you may do with the model." }, "closed-model": { "label": "dresses up … as open", "note": "Closed parts often hide in training details." } } }, "zh": { "fullName": "开源洗白 / 伪开源包装", "factExplain": "把不完整开放包装成真正开源的宣传做法。", "humanExplain": "开源洗白像直播间喊全场一元：点进去才知，只卖包装盒。\n\n看模型发布和采购评估时，用来判断开放承诺能否落地。", "humanExplainDisplay": "开源洗白像直播间喊\n==全场一元==：\n点进去才知，\n==只卖包装盒==。\n\n看模型发布和采购评估时，\n用来判断开放承诺\n能否落地。", "relationsNarrative": "Open-source Model\n开源洗白常冒充真正开源，模糊开放边界。\n\nOpen Weights\n只开放权重常被包装成完整开源。\n\nModel-licenses\n许可证决定模型能否自由使用、修改和分发。\n\nClosed-source Model\n闭源模型也可能借开放话术降低戒心。", "relations": { "open-source-model": { "label": "冒充…", "note": "真正开源要开放关键部件。" }, "open-weights": { "label": "常混淆…", "note": "只给权重，不等于完整开源。" }, "ai-model-licenses": { "label": "绕开…", "note": "许可证决定使用和改造边界。" }, "closed-model": { "label": "把…包装成开放", "note": "闭源部分常藏在训练细节里。" } } } } }, { "id": "open-weights", "name": "Open weights", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "open-source-model" }, { "to": "local-llm" }, { "to": "fine-tuning" }, { "to": "on-premise-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Open weights", "factExplain": "A model whose learned weights are openly published for anyone to download and run.", "humanExplain": "Open weights is not a free sample at the mall. It is the whole cake mix box in your bag.\n\nYou can run the model on your own computer. You can also fine-tune it or study it. Companies use it for private AI setups.", "humanExplainDisplay": "Open weights is not a ==free sample== at the mall.\nIt is the ==whole cake mix box== in your bag.\n\nYou can run the model\non your own computer.\nYou can also fine-tune it\nor study it.\nCompanies use it\nfor private AI setups.", "relationsNarrative": "Open-source-model\nOpen weights is not the same as an open-source model.\n\nLocal-LLM\nA Local-LLM usually needs downloadable weights first.\n\nFine-tuning\nFine-tuning is easier when you have the model weights.\n\nOn-premise AI\nOpen weights makes on-premise AI easier to deploy.", "relations": { "open-source-model": { "label": "is not the same as …", "note": "Open weights do not always include the training code." }, "local-llm": { "label": "helps … run", "note": "Downloaded weights make local model runs practical." }, "fine-tuning": { "label": "lets … modify", "note": "Fine-tuning needs the model weights to keep training." }, "on-premise-ai": { "label": "helps deploy …", "note": "Companies can place the model inside their own systems." } } }, "zh": { "fullName": "开放权重", "factExplain": "公开模型参数权重，允许他人下载和运行。", "humanExplain": "开放权重像超市火锅底料包：配方未必公开，但你能拿回家自己煮一锅。\n\n它让本地部署、微调和研究更方便，但许可和安全边界仍要看清。", "humanExplainDisplay": "开放权重像==超市火锅底料包==：\n配方未必公开，\n但你能拿回家==自己煮一锅==。\n\n它让本地部署、微调和研究更方便，\n但许可和安全边界仍要看清。", "relationsNarrative": "Open-source-model\n开放权重不等于开源模型，公开范围可能更窄。\n\nLocal-LLM\n本地运行模型通常要先拿到可下载的权重。\n\nFine-tuning\n只有拿到模型权重，才方便继续做微调。\n\nOn-premise AI\n开放权重让企业更容易私有化部署模型。", "relations": { "open-source-model": { "label": "不等于…", "note": "开放权重不一定连训练代码都公开。" }, "local-llm": { "label": "支撑…落地", "note": "能下载权重，才方便本地运行模型。" }, "fine-tuning": { "label": "供…继续改造", "note": "拿到权重后，才能在原模型上微调。" }, "on-premise-ai": { "label": "方便…部署", "note": "企业可把模型放进自家环境使用。" } } } } }, { "id": "openai-gym", "name": "OpenAI Gym", "layer": "L5", "sublayer": "product", "era": "2016", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "markov-decision-process" }, { "to": "deep-reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "OpenAI Gym", "factExplain": "A standard set of test worlds for reinforcement learning experiments.", "humanExplain": "OpenAI Gym is like a school gym with fixed obstacle courses. Different AI kids run them, and the stopwatch stays honest.\n\nPeople use it to train and test RL methods. It helps compare them on the same task.", "humanExplainDisplay": "OpenAI Gym is like a school gym\nwith ==fixed obstacle courses==.\nDifferent AI kids run them,\nand the ==stopwatch stays honest==.\n\nPeople use it to train and test RL methods.\nIt helps compare them on the same task.", "relationsNarrative": "RL\nOpenAI Gym is one of the most common test places for RL.\n\nMDP\nMany Gym tasks are built with states, actions, and rewards.\n\nDeep RL\nDeep RL often trains and tests in Gym environments.", "relations": { "reinforcement-learning": { "label": "hosts … experiments", "note": "Gym is a common practice field for RL." }, "markov-decision-process": { "label": "turns … into environments", "note": "Many Gym tasks use states, actions, and rewards." }, "deep-reinforcement-learning": { "label": "supports … training", "note": "Deep RL often trains and tests in Gym." } } }, "zh": { "fullName": "强化学习训练环境库", "factExplain": "一个用于强化学习实验的标准化环境工具库。", "humanExplain": "它把强化学习实验，整成统一赛道的电竞地图：英雄能换，规则、地形和计分先定死。\n\n常用来训练和测试强化学习算法，方便不同方法在同一任务里公平比较。", "humanExplainDisplay": "它把强化学习实验，\n整成统一赛道的\n==电竞地图==：\n英雄能换，\n规则、地形和\n==计分先定死==。\n\n常用来训练和测试\n强化学习算法，\n方便不同方法在同一任务里\n公平比较。", "relationsNarrative": "Reinforcement-learning\n它是强化学习里最常见的实验环境之一。\n\nMarkov-decision-process\nGym 里的很多任务按状态、动作、奖励来组织。\n\nDeep-reinforcement-learning\n深度强化学习常在它提供的环境中训练和测试。", "relations": { "reinforcement-learning": { "label": "服务…实验", "note": "它是强化学习最常见练习场之一。" }, "markov-decision-process": { "label": "把…做成环境", "note": "很多任务按状态动作奖励来设计。" }, "deep-reinforcement-learning": { "label": "支撑…训练", "note": "深度强化学习常拿它做基准实验。" } } } } }, { "id": "openpangu-2-0-flash", "name": "OpenPangu-2.0-Flash", "layer": "L3", "era": "2025", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "open-source-model" }, { "to": "tokens-per-second" } ], "track": "today", "seo": { "en": { "title": "What Is OpenPangu 2.0 Flash? Huawei's Fast Open-Source LLM", "description": "Huawei's open-source Pangu model built for speed: low-latency, low-cost generation for high-traffic chat and support. A plain-English look at what it is and where it fits." }, "zh": { "title": "OpenPangu 2.0 Flash 是什么?华为开源的高速大模型,一文看懂 — AI Rookies", "description": "面向高并发聊天与客服:低延迟、低成本,像早高峰的地铁快闸机,人流不断也不堵。人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "OpenPangu 2.0 Flash", "factExplain": "An open-source LLM built for fast text generation.", "humanExplain": "OpenPangu-2.0-Flash is the drive-thru lane with the extra headset. The line keeps moving, and nobody honks at the speaker.\n\nIt suits busy chat apps and customer support. Teams use it when they need low delay and lower cost.", "humanExplainDisplay": "OpenPangu-2.0-Flash is the ==drive-thru lane==\nwith the extra headset.\nThe line ==keeps moving==,\nand nobody honks at the speaker.\n\nIt suits busy chat apps\nand customer support.\nTeams use it when they need\nlow delay and lower cost.", "relationsNarrative": "LLM\nOpenPangu-2.0-Flash is part of the Pangu LLM family for chat and text generation.\n\nOpen-source-model\nIts open release helps developers try it, run it, and change it.\n\nTPS\nThe Flash version puts response speed front and center.", "relations": { "llm": { "label": "is a type of …", "note": "It is an LLM for writing and generating text." }, "open-source-model": { "label": "follows the … path", "note": "Its open release makes local testing and changes easier." }, "tokens-per-second": { "label": "pushes higher …", "note": "The Flash version focuses on faster text output." } } }, "zh": { "fullName": "盘古开源 2.0 闪速版", "factExplain": "面向高速推理的开源大语言模型，华为盘古系列。", "humanExplain": "OpenPangu-2.0-Flash 是早高峰地铁快闸机：人流不断，回复别堵在站口。\n\n适合高并发聊天和客服，用低延迟、低成本上线模型。", "humanExplainDisplay": "OpenPangu-2.0-Flash 是\n==早高峰地铁快闸机==：\n人流不断，\n回复==别堵在站口==。\n\n适合高并发聊天和客服，\n用低延迟、低成本\n上线模型。", "relationsNarrative": "LLM\n它是盘古系大语言模型的一支，面向对话与生成。\n\nOpen-source Model\n开放发布让开发者更容易试用、部署和改造。\n\nTPS\nFlash 版本把响应速度放在很显眼的位置。", "relations": { "llm": { "label": "属于…", "note": "它是面向文本生成的大语言模型。" }, "open-source-model": { "label": "采用…路线", "note": "开放模型便于本地试用和改造。" }, "tokens-per-second": { "label": "追求更高…", "note": "Flash 版本强调更快生成速度。" } } } } }, { "id": "optical-flow", "name": "Optical Flow", "layer": "L4", "era": "1981", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "slam" }, { "to": "assisted-driving-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Optical Flow", "factExplain": "A technique for estimating pixel motion between nearby video frames.", "humanExplain": "Optical flow is like a sports replay covered in tiny arrows. Even the hot dog wrapper gets a direction sign.\n\nIt shows how the picture moves from one frame to the next. Cars and robot cameras use it to track what moves.", "humanExplainDisplay": "Optical flow is like a ==sports replay==\ncovered in ==tiny arrows==.\nEven the hot dog wrapper\ngets a direction sign.\n\nIt shows how the picture moves\nfrom one frame to the next.\nCars and robot cameras use it\nto track what moves.", "relationsNarrative": "Computer Vision\nOptical flow gives Computer Vision a basic clue for motion in video.\n\nSLAM\nSLAM can use optical flow to track image changes and estimate pose.\n\nDriving AI\nDriving AI uses optical flow to judge how cars and people move.", "relations": { "computer-vision": { "label": "gives motion clues to …", "note": "Optical flow helps vision systems understand motion in video." }, "slam": { "label": "feeds motion estimates to …", "note": "Optical flow helps SLAM track image changes and estimate camera pose." }, "assisted-driving-ai": { "label": "helps … track motion", "note": "Optical flow helps cars judge moving vehicles and people." } } }, "zh": { "fullName": "光流", "factExplain": "估计相邻视频帧中像素运动的技术。", "humanExplain": "光流像足球回放的战术线：每个像素往哪跑，连假动作都被小箭头盯住。\n\n用于自动驾驶、视频理解和定位，帮 AI 看懂运动。", "humanExplainDisplay": "光流像足球回放的\n==战术线==：\n每个像素往哪跑，\n连假动作都被==小箭头==盯住。\n\n用于自动驾驶、视频理解和定位，\n帮 AI，\n看懂运动。", "relationsNarrative": "Computer Vision\n光流是计算机视觉理解视频运动的基础线索。\n\nSLAM\nSLAM 可用光流跟踪画面变化，辅助估计位姿。\n\nDriving AI\n自动驾驶用它判断车辆、行人的运动趋势。", "relations": { "computer-vision": { "label": "提供…运动线索", "note": "光流让视觉系统理解画面运动。" }, "slam": { "label": "为…提供运动估计", "note": "相机位姿可借光流跟踪变化。" }, "assisted-driving-ai": { "label": "帮助…追踪动态", "note": "车辆行人移动可由光流辅助判断。" } } } } }, { "id": "optimization", "name": "Optimization", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "gradient-descent" }, { "to": "sgd" }, { "to": "adam" }, { "to": "backpropagation" } ], "track": "history", "i18n": { "en": { "fullName": "Optimization", "factExplain": "The process of nudging model settings toward better results during training.", "humanExplain": "Optimization is like a fussy hotel shower knob. You keep tiny-turning it until the water stops attacking you.\n\nIn model training, it keeps nudging the model’s settings. Each nudge aims for better answers.", "humanExplainDisplay": "Optimization is like a ==fussy hotel shower knob==.\nYou keep ==tiny-turning it==\nuntil the water stops attacking you.\n\nIn model training,\nit keeps nudging the model’s settings.\nEach nudge aims for better answers.", "relationsNarrative": "Gradient Descent\nGradient Descent is the classic way to do optimization.\n\nSGD\nSGD updates settings with small batches of data.\n\nAdam\nAdam adjusts step size, so optimization often moves faster.\n\nBackpropagation\nBackprop gives the direction for each optimization step.", "relations": { "gradient-descent": { "label": "often uses …", "note": "Gradient Descent is the classic way to optimize a model." }, "sgd": { "label": "has common version …", "note": "SGD updates settings with small batches of data." }, "adam": { "label": "has faster version …", "note": "Adam adjusts step size, so training often improves faster." }, "backpropagation": { "label": "gets direction from …", "note": "Backprop tells optimization which way to move the settings." } } }, "zh": { "fullName": "优化", "factExplain": "让模型参数朝更优目标不断调整的过程。", "humanExplain": "像煎饼摊师傅调面糊火候：稀一点焦了，厚一点夹生，得一遍遍试到那张皮最对劲。\n\n它贯穿模型训练，负责把参数一步步调到更好的位置。", "humanExplainDisplay": "像煎饼摊师傅调==面糊火候==：\n稀一点焦了，厚一点夹生，\n得一遍遍试到那张皮==最对劲==。\n\n它贯穿模型训练，\n负责把参数一步步\n调到更好的位置。", "relationsNarrative": "Gradient Descent\n梯度下降是优化最经典、最基础的实现方式。\n\nSGD\nSGD 是优化的常见做法，用小批量数据更新参数。\n\nAdam\nAdam 是常用优化器，收敛通常比基础方法更快。\n\nBackpropagation\n反向传播先算出梯度，优化再据此更新参数。", "relations": { "gradient-descent": { "label": "常用…实现", "note": "梯度下降是最经典的优化方法。" }, "sgd": { "label": "有经典变体…", "note": "SGD 是优化里最常见的基础做法。" }, "adam": { "label": "有高效变体…", "note": "Adam 用自适应步长加快收敛。" }, "backpropagation": { "label": "靠…提供方向", "note": "反向传播告诉优化该往哪调参数。" } } } } }, { "id": "options-framework", "name": "Options Framework", "layer": "L2", "era": "1999", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "markov-decision-process" } ], "track": "history", "i18n": { "en": { "fullName": "Options Framework", "factExplain": "A reinforcement learning framework where one choice can run a reusable multi-step action.", "humanExplain": "Options Framework is like hitting one combo button in a video game. Your fighter does punch-kick-spin, while your thumb takes a tiny vacation.\n\nIt is used in HRL. It lets long tasks reuse multi-step skills.", "humanExplainDisplay": "Options Framework is like hitting one ==combo button==\nin a video game.\nYour fighter does ==punch-kick-spin==,\nwhile your thumb takes a tiny vacation.\n\nIt is used in HRL.\nIt lets long tasks reuse multi-step skills.", "relationsNarrative": "HRL\nThe Options Framework is a classic tool for HRL.\n\nRL\nIt extends RL actions into multi-step skills.\n\nMDP\nOptions usually define start rules, behavior rules, and stop rules on an MDP.", "relations": { "reinforcement-learning": { "label": "extends … actions", "note": "It turns one-step actions into multi-step skills." }, "markov-decision-process": { "label": "is defined on …", "note": "Start rules, behavior rules, and stop rules use the state." } } }, "zh": { "fullName": "选项框架", "factExplain": "用“选项”表示可持续多步动作的强化学习框架。", "humanExplain": "选项框架像游戏一键连招：不用每下平A都点，按个技能就打一套。\n\n用于层级强化学习，让长任务复用多步技能。", "humanExplainDisplay": "选项框架像游戏一键连招：\n不用每下==平A都点==，\n按个技能就==打一套==。\n\n用于层级强化学习，\n让长任务复用多步技能。", "relationsNarrative": "Hierarchical Reinforcement Learning\n选项框架是层级强化学习的经典形式化工具。\n\nReinforcement Learning\n它把强化学习里的动作扩展成多步技能。\n\nMDP\n选项通常在 MDP 上定义启动、策略和终止。", "relations": { "reinforcement-learning": { "label": "扩展…动作", "note": "它把单步动作升级为多步技能。" }, "markov-decision-process": { "label": "定义在…上", "note": "启动、策略、终止都依赖环境状态。" } } } } }, { "id": "outlier-detection", "name": "Outlier Detection", "layer": "L4", "era": "1969", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "unsupervised-learning" }, { "to": "clustering" }, { "to": "dbscan" }, { "to": "ai-anti-cheat" } ], "track": "history", "i18n": { "en": { "fullName": "Outlier Detection", "factExplain": "A way to spot data points that are far from what is normal.", "humanExplain": "Outlier detection is a cafeteria monitor with a raised eyebrow. Most kids carry trays. One kid rolls in with a ketchup cart.\n\nYou meet it in fraud checks, server alerts, and factory checks. It flags odd cases first, so a person can review them.", "humanExplainDisplay": "Outlier detection is a ==cafeteria monitor==\nwith a raised eyebrow.\nMost kids carry trays.\nOne kid rolls in\nwith a ==ketchup cart==.\n\nYou meet it in fraud checks,\nserver alerts,\nand factory checks.\nIt flags odd cases first,\nso a person can review them.", "relationsNarrative": "Unsupervised Learning\nOutlier detection often finds strange samples in data with no labels.\n\nClustering\nClustering helps find points far from the group.\n\nDBSCAN\nDBSCAN marks lonely low-density points as noise.\n\nAI Anti-Cheat\nAI Anti-Cheat uses outlier detection to spot suspicious behavior.", "relations": { "unsupervised-learning": { "label": "often uses …", "note": "It can find odd samples even without labels." }, "clustering": { "label": "uses … to spot loners", "note": "Outliers are often data points far from any group." }, "dbscan": { "label": "marks noise with …", "note": "DBSCAN labels lonely, low-density points as noise." }, "ai-anti-cheat": { "label": "helps … catch odd behavior", "note": "Odd behavior can be a clue for cheating." } } }, "zh": { "fullName": "异常值检测 / 离群点检测", "factExplain": "识别数据中显著偏离常态样本的方法。", "humanExplain": "异常值检测就是小区群防诈：大家正常进出，谁半夜扛冰箱翻墙，先拉黑再盘问。\n\n用于风控、运维、质检，先挑可疑样本给人复核。", "humanExplainDisplay": "异常值检测就是==小区群防诈==：\n大家正常进出，\n谁半夜==扛冰箱翻墙==，\n先拉黑再盘问。\n\n用于风控、运维、质检，\n先挑可疑样本，\n给人复核。", "relationsNarrative": "Unsupervised Learning\n异常值检测常在无标签数据里寻找不正常样本。\n\nClustering\n聚类能帮助发现远离群体的离群点。\n\nDBSCAN\nDBSCAN 会把密度太低的孤立点标成噪声。\n\nAI Anti-Cheat\n反作弊会用异常值检测发现可疑行为。", "relations": { "unsupervised-learning": { "label": "常用…完成", "note": "没标签时也能找不正常样本。" }, "clustering": { "label": "借…发现偏离", "note": "离群点常是不合群的数据。" }, "dbscan": { "label": "用…标噪声", "note": "DBSCAN 会把孤立点标成噪声。" }, "ai-anti-cheat": { "label": "帮…抓异常", "note": "异常行为是作弊线索之一。" } } } } }, { "id": "overparameterization", "name": "Overparameterization", "layer": "L2", "era": "2018", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "regularization" }, { "to": "deep-learning" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Overparameterization", "factExplain": "A model has more parameters than it needs to fit the training data.", "humanExplain": "Overparameterization is a giant box of crayons for one tiny drawing. The page is tiny, but a good kid can still make the dog look great.\n\nBig deep models use this a lot to train more easily. But they need good data and regularization.", "humanExplainDisplay": "Overparameterization is a ==giant box of crayons==\nfor one ==tiny drawing==.\nThe page is tiny,\nbut a good kid can still\nmake the dog look great.\n\nBig deep models use this a lot\nto train more easily.\nBut they need good data\nand regularization.", "relationsNarrative": "Parameter\nOverparameterization means the model has more parameters than it needs.\n\nRegularization\nRegularization controls the extra freedom and reduces overfitting.\n\nDeep Learning\nDeep Learning often uses huge parameter counts to make training easier.\n\nBias-Variance Tradeoff\nIt makes the old \"more parameters means overfitting\" rule less simple.", "relations": { "parameter": { "label": "adds more …", "note": "More parameters make overparameterization more likely." }, "regularization": { "label": "is held in check by …", "note": "Regularization helps a big model avoid memorizing the training set." }, "deep-learning": { "label": "is common in …", "note": "Deep networks often use many parameters to make training easier." }, "bias-variance-tradeoff": { "label": "complicates …", "note": "It makes \"more parameters means overfitting\" too simple." } } }, "zh": { "fullName": "过度参数化", "factExplain": "模型参数数量超过拟合训练数据所需的现象。", "humanExplain": "过度参数化像用 500 色蜡笔画简笔画小狗：颜色多到用不完，挑起来反而顺手。\n\n常见于深度大模型，让训练更顺，但更依赖数据与正则化。", "humanExplainDisplay": "过度参数化像用\n==500 色蜡笔==画简笔画小狗：\n颜色多到用不完，\n挑起来==反而顺手==。\n\n常见于深度大模型，\n让训练更顺，\n但更依赖数据与正则化。", "relationsNarrative": "Parameter\n过度参数化指参数数量超过拟合所需。\n\nRegularization\n正则化约束多余自由度，减少过拟合。\n\nDeep Learning\n深度学习常用海量参数提升可训练性。\n\nBias-Variance Tradeoff\n它让“参数多必过拟合”的直觉变复杂。", "relations": { "parameter": { "label": "增加…数量", "note": "参数越多，越容易进入过度参数化。" }, "regularization": { "label": "靠…约束", "note": "正则化帮大模型别把训练集背死。" }, "deep-learning": { "label": "常见于…", "note": "深度网络常用大量参数换可训练性。" }, "bias-variance-tradeoff": { "label": "挑战…直觉", "note": "它让传统过拟合直觉不再够用。" } } } } }, { "id": "pac-learning", "name": "PAC", "layer": "L2", "era": "1984", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "statistical-learning-theory" }, { "to": "supervised-learning" }, { "to": "empirical-risk-minimization" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Probably Approximately Correct Learning", "factExplain": "A framework for saying when a model learns well enough, with probability.", "humanExplain": "PAC learning is like a frozen pizza brand. Not perfect, but it should not ruin Friday night often.\n\nIt is a basic ruler for learning theory. It checks if you have enough examples. It asks if the model will work on new cases.", "humanExplainDisplay": "PAC learning is like a ==frozen pizza brand==.\nNot perfect,\nbut it should not ==ruin Friday night== often.\n\nIt is a basic ruler for learning theory.\nIt checks if you have enough examples.\nIt asks if the model will work on new cases.", "relationsNarrative": "SLT\nPAC is one classic way SLT defines learning.\n\nSupervised Learning\nPAC often studies when Supervised Learning can learn from examples.\n\nERM\nPAC explains why low training error may not mean good new-case performance.\n\nBias-Variance Tradeoff\nBoth ask how model complexity affects new-case performance.", "relations": { "statistical-learning-theory": { "label": "belongs to …", "note": "PAC is a classic core problem in SLT." }, "supervised-learning": { "label": "often analyzes …", "note": "PAC often asks when Supervised Learning can really learn." }, "empirical-risk-minimization": { "label": "warns about …", "note": "Good training scores do not always mean good new-case scores." }, "bias-variance-tradeoff": { "label": "links to …", "note": "Both study how model complexity affects new-case performance." } } }, "zh": { "fullName": "Probably Approximately Correct Learning／概率近似正确学习", "factExplain": "用概率方式定义“学得够好”的学习框架。", "humanExplain": "PAC 学习像食堂打饭：不求每勺都惊艳，但大多数时候别翻车，长期吃着稳，这标准就算靠谱。\n\n它用来分析样本够不够、模型能否泛化，是学习理论的基础尺子。", "humanExplainDisplay": "PAC 学习像食堂打饭：\n不求每勺都==惊艳==，\n但大多数时候别翻车，\n长期吃着稳，\n这标准就算\n==靠谱==。\n\n它用来分析\n样本够不够、模型能否泛化，\n是学习理论的基础尺子。", "relationsNarrative": "Statistical-learning-theory\n它是统计学习理论里最经典的学习定义之一。\n\nSupervised-learning\nPAC 学习常用来分析监督学习何时能学会。\n\nEmpirical-risk-minimization\n它解释为什么训练集表现好，未必代表真能泛化。\n\nBias-variance-tradeoff\n两者都关心模型复杂度会怎样影响泛化。", "relations": { "statistical-learning-theory": { "label": "属于…框架", "note": "它是统计学习理论的经典核心问题。" }, "supervised-learning": { "label": "常用于分析…", "note": "最常用来讨论监督学习能否学会。" }, "empirical-risk-minimization": { "label": "给…做担保", "note": "说明经验上学得好不等于真会泛化。" }, "bias-variance-tradeoff": { "label": "关联…取舍", "note": "都在讨论模型复杂度与泛化表现。" } } } } }, { "id": "pagerank", "name": "PageRank", "layer": "L2", "era": "1998", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "information-retrieval" }, { "to": "graph-search" }, { "to": "bm25" } ], "track": "history", "i18n": { "en": { "fullName": "PageRank", "factExplain": "An algorithm that ranks web pages by the links between them.", "humanExplain": "PageRank is a popularity contest for web pages. A nod from the trusted kid beats ten random high-fives.\n\nIt helped early search engines rank pages. It also helps find important spots in a graph.", "humanExplainDisplay": "PageRank is a ==popularity contest== for web pages.\nA nod from the ==trusted kid==\nbeats ten random high-fives.\n\nIt helped early search engines rank pages.\nIt also helps find important spots in a graph.", "relationsNarrative": "IR\nPageRank was a key early tool for search ranking.\n\nGraph Search\nPageRank treats web links as a graph and moves importance along links.\n\nBM25\nBM25 checks keyword matches, and PageRank adds link authority.", "relations": { "information-retrieval": { "label": "helped power … ranking", "note": "Early search used PageRank to judge a page's weight." }, "graph-search": { "label": "uses … ideas", "note": "Web links can be treated as one giant graph." }, "bm25": { "label": "adds link authority to …", "note": "BM25 checks word matches. PageRank checks link authority." } } }, "zh": { "fullName": "网页排名算法", "factExplain": "按链接关系评估网页重要性的排序算法。", "humanExplain": "PageRank评网页，走班级评优路子：不看谁嗓门大，看谁被靠谱同学提名。\n\n它曾支撑网页搜索排序，也用于图上的重要性计算。", "humanExplainDisplay": "PageRank评网页，\n走==班级评优==路子：\n不看谁==嗓门大==，\n看谁被靠谱同学提名。\n\n它曾支撑网页搜索排序，\n也用于图上的重要性计算。", "relationsNarrative": "Information Retrieval\nPageRank 是早期搜索排序的关键算法之一。\n\nGraph Search\n它把网页链接看作图，让重要性沿边传递。\n\nBM25\nBM25 看关键词匹配，它补上链接权威性。", "relations": { "information-retrieval": { "label": "支撑…排序", "note": "搜索排序曾靠它判断网页分量。" }, "graph-search": { "label": "基于…思想", "note": "网页链接被当成一张巨大图。" }, "bm25": { "label": "补充…排序", "note": "一个看链接权重，一个看词匹配。" } } } } }, { "id": "parameter-efficient-fine-tuning", "name": "PEFT", "layer": "L2", "era": "2019", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "fine-tuning" }, { "to": "lora" }, { "to": "base-model" } ], "track": "history", "i18n": { "en": { "fullName": "Parameter-Efficient Fine-Tuning", "factExplain": "A fine-tuning method that updates only a few model parameters.", "humanExplain": "PEFT is like a clip-on basket for your old bike. The bike stays the same. Now it carries snacks.\n\nTeams use it for private custom AI, with less memory and cost. But it still hits the base model’s limits.", "humanExplainDisplay": "PEFT is like a ==clip-on basket==\nfor your old bike.\nThe bike stays the same.\nNow it ==carries snacks==.\n\nTeams use it for private custom AI,\nwith less memory and cost.\nBut it still hits\nthe base model’s limits.", "relationsNarrative": "Fine-tuning\nPEFT is fine-tuning with fewer trainable parameters.\n\nLoRA\nLoRA is a common PEFT method for small low-rank changes.\n\nBase model\nPEFT usually freezes most of the base model and trains only small added parts.", "relations": { "fine-tuning": { "label": "is a cheaper version of …", "note": "PEFT keeps the fine-tuning goal but changes fewer parameters." }, "lora": { "label": "is often done with …", "note": "LoRA is the most common way to do PEFT." }, "base-model": { "label": "freezes most of …", "note": "PEFT usually leaves the base model mostly unchanged." } } }, "zh": { "fullName": "参数高效微调", "factExplain": "只更新少量参数来适配任务的微调方法。", "humanExplain": "PEFT 像给旧西装缝个新口袋：不拆整件衣服，也能装下新任务。\n\n常用于私有定制，省显存省成本，但受底座上限限制。", "humanExplainDisplay": "PEFT 像给旧西装\n缝个==新口袋==：\n不拆整件衣服，\n也能装下新任务。\n\n常用于私有定制，\n省显存省成本，\n但受底座上限限制。", "relationsNarrative": "Fine-tuning\nPEFT 是微调的省参数做法，只更新少量权重。\n\nLoRA\nLoRA 是最常见的 PEFT 方法之一，训练低秩增量。\n\nBase Model\nPEFT 通常冻结底座主体，只加少量可训练部分。", "relations": { "fine-tuning": { "label": "是…的省钱版", "note": "它少改参数，保留微调目标。" }, "lora": { "label": "常用…实现", "note": "LoRA 是最常见的参数高效方案。" }, "base-model": { "label": "冻结…主体", "note": "底座模型大多不动，只学小模块。" } } } } }, { "id": "parameter-server", "name": "Parameter Server", "layer": "L5", "sublayer": "architecture", "era": "2010", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "data-parallelism" }, { "to": "sgd" }, { "to": "distributed-computing" } ], "track": "history", "i18n": { "en": { "fullName": "Parameter Server", "factExplain": "A setup that keeps model parameters in one place during distributed training.", "humanExplain": "Parameter Server is the lunch lady with the only real recipe card. The kitchen crew sends tweaks, and she hands back the new card.\n\nIn big AI training, many machines work at once. They send weight updates to the server, then get the latest weights back.", "humanExplainDisplay": "Parameter Server is the ==lunch lady==\nwith the only ==real recipe card==.\nThe kitchen crew sends tweaks,\nand she hands back the new card.\n\nIn big AI training,\nmany machines work at once.\nThey send weight updates to the server,\nthen get the latest weights back.", "relationsNarrative": "Parameter\nA Parameter Server keeps Parameters in one central place.\n\nData Parallelism\nData Parallelism lets many machines work at once, then send updates back.\n\nSGD\nA Parameter Server often coordinates SGD weight updates.\n\nDistributed Computing\nA Parameter Server is a classic setup in Distributed Computing.", "relations": { "parameter": { "label": "keeps … in one place", "note": "It stores and updates model parameters in one shared spot." }, "data-parallelism": { "label": "supports … training", "note": "Many machines compute gradients, then send updates back." }, "sgd": { "label": "coordinates … updates", "note": "It often gathers gradients and updates weights for SGD." }, "distributed-computing": { "label": "is a … pattern", "note": "It is a classic coordination setup for distributed training." } } }, "zh": { "fullName": "参数服务器", "factExplain": "分布式训练中集中管理模型参数的架构。", "humanExplain": "参数服务器像乐队总谱：乐手各练各的，改动都回总谱汇总，再照最新版一起演。\n\n用于大规模训练，集中同步权重更新。", "humanExplainDisplay": "参数服务器像\n==乐队总谱==：\n乐手各练各的，改动\n都回总谱汇总，\n再照==最新版一起演==。\n\n用于大规模训练，\n集中同步权重更新。", "relationsNarrative": "Parameter\n参数服务器集中保存并更新模型参数。\n\nData Parallelism\n数据并行让多台机器同时算，再回传更新。\n\nSGD\nSGD 的梯度更新常由它统一协调。\n\nDistributed Computing\n它是分布式机器学习里的经典架构。", "relations": { "parameter": { "label": "集中管理…", "note": "它把模型参数放在统一位置更新。" }, "data-parallelism": { "label": "支撑…训练", "note": "多台机器算梯度，再回传更新。" }, "sgd": { "label": "协调…更新", "note": "常接收各节点梯度来更新权重。" }, "distributed-computing": { "label": "属于…架构", "note": "它是分布式训练的经典协调方式。" } } } } }, { "id": "parameter", "name": "Parameter", "layer": "L2", "era": "2010s", "publishedAt": "2026-05-23T08:45:00Z", "relations": [ { "to": "neural-network" }, { "to": "pretraining" }, { "to": "foundation-model" }, { "to": "scaling-law" } ], "i18n": { "en": { "fullName": "Parameter", "factExplain": "A number inside a model learned during training and used to make outputs.", "humanExplain": "Parameters are the tiny knobs inside an AI. Training keeps turning them until it stops sounding like a confused parrot.\n\nMore parameters do not mean the AI is always right. They often mean more room to learn. They also mean a scarier bill.", "humanExplainDisplay": "Parameters are the ==tiny knobs== inside an AI.\nTraining keeps turning them\nuntil it stops sounding like a ==confused parrot==.\n\nMore parameters do not mean\nthe AI is always right.\nThey often mean more room to learn.\nThey also mean a scarier bill.", "relationsNarrative": "Neural-network\nParameter is an internal value a Neural-network learns through training.\n\nPretraining\nMore parameters often need larger Pretraining to work well.\n\nFoundation-model\nA Foundation-model's limit often depends on its parameter count.\n\nScaling-law\nScaling-law describes how more parameters can improve performance.", "relations": { "neural-network": { "label": "are core numbers in …", "note": "Parameters are the internal numbers a Neural-network learns during training." }, "pretraining": { "label": "are tuned by …", "note": "Pretraining adjusts many parameters so the model becomes useful." }, "foundation-model": { "label": "make up … at scale", "note": "A Foundation-model is built from a huge number of parameters." }, "scaling-law": { "label": "are studied by …", "note": "Scaling-law links parameter growth to model performance." } } }, "zh": { "fullName": "参数", "factExplain": "模型在训练中学到并用于计算输出的内部数值。", "humanExplain": "参数就像奶茶店配方里的糖和冰，调多少，训练完都记在小本本上。\n\n它决定模型怎么判断和生成，也常被用来粗略衡量模型规模。", "humanExplainDisplay": "参数就像==奶茶店配方里的糖和冰==，\n调多少，\n训练完都记在==小本本==上。\n\n它决定模型怎么判断和生成，\n也常被用来粗略衡量模型规模。", "relationsNarrative": "Neural-network\nParameter 是 Neural-network 通过训练学到的内部变量。\n\nPretraining\n更多 Parameter 往往需要更大规模的 Pretraining 才能发挥作用。\n\nFoundation-model\nFoundation-model 的能力上限常与 Parameter 规模相关。\n\nScaling-law\nScaling-law 描述 Parameter 增长与性能提升的关系。", "relations": { "neural-network": { "label": "是…的核心数值" }, "pretraining": { "label": "被…不断调整" }, "foundation-model": { "label": "大规模构成…" }, "scaling-law": { "label": "规模由…研究" } } } } }, { "id": "partially-observable-markov-decision-process", "name": "POMDP", "layer": "L2", "era": "1965", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "markov-decision-process" }, { "to": "belief-state" }, { "to": "reinforcement-learning" }, { "to": "hidden-markov-model" } ], "track": "history", "i18n": { "en": { "fullName": "Partially Observable Markov Decision Process", "factExplain": "A math model for making step-by-step choices when you cannot see everything.", "humanExplain": "POMDP is like playing a racing game while someone blocks half the TV. You guess the track from tiny clues, then steer anyway.\n\nRobots use it. So do self-driving cars and chat agents. It helps them watch a bit, act, then watch again.", "humanExplainDisplay": "POMDP is like playing a racing game\nwhile someone ==blocks half the TV==.\nYou guess the track from ==tiny clues==,\nthen steer anyway.\n\nRobots use it.\nSo do self-driving cars and chat agents.\nIt helps them watch a bit,\nact,\nthen watch again.", "relationsNarrative": "MDP\nA POMDP is an MDP with part of the world hidden.\n\nBelief State\nA POMDP uses a Belief State when the real state is unclear.\n\nRL\nMany real RL tasks fit a POMDP better than a fully visible world.\n\nHMM\nA POMDP uses the hidden-state idea from an HMM, then adds actions.", "relations": { "markov-decision-process": { "label": "adds blind spots to …", "note": "A POMDP takes an MDP and hides part of the world." }, "belief-state": { "label": "decides using …", "note": "A Belief State is the agent’s best guess when it cannot see enough." }, "reinforcement-learning": { "label": "often frames … tasks", "note": "Many real RL jobs have missing information, so they fit POMDPs." }, "hidden-markov-model": { "label": "borrows hidden states from …", "note": "Like an HMM, it tracks hidden state, but it also chooses actions." } } }, "zh": { "fullName": "部分可观测马尔可夫决策过程", "factExplain": "在信息不完整时做序列决策的数学框架。", "humanExplain": "POMDP像相亲时隔着滤镜聊天：你看不清对方全貌，只能凭几句线索，边猜边决定要不要继续。\n\n它常用于机器人、自动驾驶和对话代理，适合边观察边行动。", "humanExplainDisplay": "POMDP像相亲时隔着滤镜聊天：\n你看不清对方==全貌==，\n只能凭几句==线索==，\n边猜边决定要不要继续。\n\n它常用于机器人、自动驾驶\n和对话代理，\n适合边观察边行动。", "relationsNarrative": "Markov Decision Process\n它是在 MDP 基础上，加上“看不全环境”的设定。\n\nBelief State\n看不清真实状态时，常靠信念状态来做决策。\n\nReinforcement Learning\n很多现实强化学习任务，本质上都可建模成它。\n\nHidden Markov Model\n它借了 HMM 的隐藏状态思路，但还要负责选动作。", "relations": { "markov-decision-process": { "label": "是在…上加遮挡", "note": "它把全可观测决策扩展到信息不全场景。" }, "belief-state": { "label": "靠…做判断", "note": "看不全环境时，要靠信念状态补信息。" }, "reinforcement-learning": { "label": "常作为…问题设定", "note": "很多现实强化学习任务都更像它。" }, "hidden-markov-model": { "label": "结合…的隐藏状态思路", "note": "它既要推断状态，也要选择动作。" } } } } }, { "id": "particle-filter", "name": "Particle Filter", "layer": "L2", "era": "1993", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "kalman-filter" }, { "to": "hidden-markov-model" }, { "to": "markov-chain-monte-carlo" }, { "to": "slam" } ], "track": "history", "i18n": { "en": { "fullName": "Particle Filter", "factExplain": "A method for tracking a hidden state with many random guesses.", "humanExplain": "Particle Filter is like finding your lost dog at a park. Drop guesses like tennis balls on a map, then keep the ones near fresh paw prints.\n\nIt helps robots and apps track position. It can still follow the state through messy sensor noise.", "humanExplainDisplay": "Particle Filter is like finding your ==lost dog== at a park.\nDrop guesses like ==tennis balls on a map==,\nthen keep the ones near fresh paw prints.\n\nIt helps robots and apps track position.\nIt can still follow the state\nthrough messy sensor noise.", "relationsNarrative": "Kalman Filter\nBoth estimate state, but Particle Filter handles nonlinear paths better.\n\nHMM\nParticle Filter can use samples to estimate the hidden state in an HMM.\n\nMCMC\nBoth use sampling, but Particle Filter is built for live tracking.\n\nSLAM\nSLAM can use Particle Filter to locate a robot through noisy motion and sensor data.", "relations": { "kalman-filter": { "label": "also estimates state like …", "note": "Particle Filter handles more nonlinear cases than Kalman Filter." }, "hidden-markov-model": { "label": "estimates hidden state in …", "note": "It uses samples to track a hidden state over time." }, "markov-chain-monte-carlo": { "label": "also uses sampling like …", "note": "Particle Filter tracks online, while MCMC often samples a fixed target." }, "slam": { "label": "helps … locate itself", "note": "Robots can use it to estimate their position while moving." } } }, "zh": { "fullName": "粒子滤波", "factExplain": "用大量随机样本递推估计隐状态的方法。", "humanExplain": "粒子滤波像商场寻娃：先派一圈亲戚瞎猜，线索对上的留下。\n\n用于定位和跟踪，噪声再大也能追状态。", "humanExplainDisplay": "粒子滤波像商场寻娃：\n先派==一圈亲戚==瞎猜，\n线索对上的==留下==。\n\n用于定位和跟踪，\n噪声再大，\n也能追状态。", "relationsNarrative": "Kalman Filter\n两者都做状态估计，粒子滤波更能处理非线性和多峰分布。\n\nHidden Markov Model\n它常在 HMM 中用采样近似隐状态的后验分布。\n\nMCMC\n它和 MCMC 都靠采样近似概率分布，但更偏在线追踪。\n\nSLAM\nSLAM 可用粒子滤波在运动和观测噪声中定位。", "relations": { "kalman-filter": { "label": "同做状态估计", "note": "粒子滤波适合更非线性的场景。" }, "hidden-markov-model": { "label": "估计…的隐状态", "note": "它用样本追踪随时间变化的隐状态。" }, "markov-chain-monte-carlo": { "label": "同属采样方法", "note": "它偏在线递推，MCMC多用于静态采样。" }, "slam": { "label": "支撑…定位", "note": "机器人可用它边走边估计自身位置。" } } } } }, { "id": "pddl", "name": "PDDL", "layer": "L2", "era": "1998", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "automated-planning" }, { "to": "strips" }, { "to": "knowledge-representation" }, { "to": "symbolic-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Planning Domain Definition Language", "factExplain": "A planning language for writing states, actions, and goals.", "humanExplain": "PDDL is like a board game rule sheet for a robot. It shows the start, legal moves, and how to win.\n\nYou meet it in robots, games, and delivery planning. A planner reads it and finds a valid plan.", "humanExplainDisplay": "PDDL is like a ==board game rule sheet==\nfor a robot.\nIt shows the start,\n==legal moves==,\nand how to win.\n\nYou meet it in robots, games,\nand delivery planning.\nA planner reads it\nand finds a valid plan.", "relationsNarrative": "Planning\nPDDL describes the world and the goal for a planner.\n\nSTRIPS\nPDDL inherits and extends the STRIPS way of writing actions.\n\nKR\nPDDL writes states, actions, and goals as symbols a system can reason with.\n\nSymbolic AI\nPDDL belongs to the symbolic AI tradition for planning.", "relations": { "automated-planning": { "label": "describes … problems", "note": "PDDL is often the input a planner reads." }, "strips": { "label": "extends … actions", "note": "PDDL keeps the STRIPS idea of actions changing the world." }, "knowledge-representation": { "label": "writes the world with …", "note": "PDDL writes states, actions, and goals as symbols." }, "symbolic-ai": { "label": "belongs to …", "note": "PDDL is a classic language from symbolic planning." } } }, "zh": { "fullName": "规划领域定义语言（Planning Domain Definition Language）", "factExplain": "描述状态、动作与目标的规划语言。", "humanExplain": "PDDL 像给机器人写旅行攻略：起点、车票、景点顺序全写，别只说“玩开心”。\n\n用于机器人、游戏、物流规划，让求解器按规则找路。", "humanExplainDisplay": "PDDL 像给机器人写==旅行攻略==：\n==起点、车票、景点顺序==全写，\n别只说“玩开心”。\n\n用于机器人、游戏、物流规划，\n让求解器按规则，\n找路。", "relationsNarrative": "Automated Planning\nPDDL 常作为规划器的输入格式，描述问题和领域。\n\nSTRIPS\nPDDL 继承并扩展了 STRIPS 的动作表示。\n\nKnowledge Representation\nPDDL 把状态、动作和目标写成可推理的符号。\n\nSymbolic AI\nPDDL 属于符号 AI 的规划语言传统。", "relations": { "automated-planning": { "label": "描述…问题", "note": "PDDL 常作为规划器的输入格式。" }, "strips": { "label": "扩展…表达", "note": "PDDL 继承了 STRIPS 的动作思想。" }, "knowledge-representation": { "label": "用…表达世界", "note": "它把状态、动作和目标写成符号。" }, "symbolic-ai": { "label": "属于…传统", "note": "PDDL 是典型的符号式规划语言。" } } } } }, { "id": "penn-treebank", "name": "Penn Treebank", "layer": "L4", "era": "1993", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "syntactic-parsing" }, { "to": "sequence-labeling" }, { "to": "language-modeling" } ], "track": "history", "i18n": { "en": { "fullName": "Penn Treebank", "factExplain": "A set of English text marked with word roles and sentence trees.", "humanExplain": "Penn Treebank is like a grammar teacher with a giant red pen. Every word gets a name tag, and every sentence gets a family tree.\n\nIt helps train tools for word labels and sentence structure. Older language models also used it as a scoreboard.", "humanExplainDisplay": "Penn Treebank is like a grammar teacher\nwith a ==giant red pen==.\nEvery word gets a name tag,\nand every sentence gets a ==family tree==.\n\nIt helps train tools\nfor word labels and sentence structure.\nOlder language models also used it\nas a scoreboard.", "relationsNarrative": "Syntax Parse\nPenn Treebank gives human-made syntax trees for training and testing.\n\nSequence Labeling\nIts part-of-speech tags help train sequence labeling models.\n\nLM\nEarly language models used Penn Treebank as a perplexity benchmark.", "relations": { "syntactic-parsing": { "label": "trains …", "note": "Penn Treebank is a classic dataset for syntax parsing." }, "sequence-labeling": { "label": "supports …", "note": "Its part-of-speech tags are used for sequence labeling." }, "language-modeling": { "label": "benchmarks …", "note": "Early language models used it to test perplexity." } } }, "zh": { "fullName": "宾州树库", "factExplain": "带词性与句法树标注的英语语料库。", "humanExplain": "宾州树库像英语句子的施工图：哪个词当梁柱，短语怎么搭，都画成树。\n\n用于词性标注、句法解析，也给语言模型做基准。", "humanExplainDisplay": "宾州树库像英语句子的\n==施工图==：\n哪个词当梁柱，\n短语怎么搭都==画成树==。\n\n用于词性标注、句法解析，\n也给语言模型，\n做基准。", "relationsNarrative": "Syntax Parse\n宾州树库提供人工句法树，是经典训练与评测集。\n\nSequence Labeling\n其中的词性标注常用于训练序列标注模型。\n\nLM\n早期语言模型常用它做困惑度基准。", "relations": { "syntactic-parsing": { "label": "提供…训练数据", "note": "它是句法解析的经典语料。" }, "sequence-labeling": { "label": "支持…标注任务", "note": "词性标注可视为序列标注。" }, "language-modeling": { "label": "作为…基准", "note": "早期语言模型常用它测困惑度。" } } } } }, { "id": "perceptron", "name": "Perceptron", "layer": "L1", "era": "1958", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "mlp" }, { "to": "backpropagation" }, { "to": "connectionism" } ], "track": "history", "i18n": { "en": { "fullName": "Perceptron", "factExplain": "The earliest one-layer AI neuron for sorting things into two groups.", "humanExplain": "A Perceptron is a roller coaster height sign. Tall enough, you ride. One hair short, enjoy the bench.\n\nIt helps explain simple classification and weights. You meet it in the first lesson on neural networks.", "humanExplainDisplay": "A Perceptron is a ==roller coaster height sign==.\nTall enough, you ride.\nOne hair short,\n==enjoy the bench==.\n\nIt helps explain simple classification and weights.\nYou meet it in the first lesson on neural networks.", "relationsNarrative": "Neural-network\nThe Perceptron was an early starting point for Neural-network.\n\nMLP\nMLP is a direct extension of the Perceptron with more layers.\n\nBackpropagation\nAfter the Perceptron, deeper networks used Backprop to learn.\n\nConnectionism\nThe Perceptron is a classic early model in Connectionism.", "relations": { "neural-network": { "label": "is a starting point for …", "note": "It helped start the later story of neural networks." }, "mlp": { "label": "grows into …", "note": "MLP adds layers to fix the limits of one layer." }, "backpropagation": { "label": "paved the way for …", "note": "Deeper networks later used Backprop to learn." }, "connectionism": { "label": "belongs to …", "note": "It models intelligence with connected neuron-like units." } } }, "zh": { "fullName": "感知机", "factExplain": "最早的单层神经元二分类模型。", "humanExplain": "像球场裁判吹哨：动作够线就算进，差一点就判掉，判断简单，但边界分得很死。\n\n常用来理解分类、权重和神经网络最早怎么工作。", "humanExplainDisplay": "像球场裁判吹哨：\n动作==够线就算进==，\n差一点就判掉，\n判断简单，但==边界分得很死==。\n\n常用来理解分类、\n权重和神经网络\n最早怎么工作。", "relationsNarrative": "Neural-network\n它是神经网络发展史上的早期起点之一。\n\nMultilayer-perceptron\n多层感知机是在它基础上的直接扩展。\n\nBackpropagation\n单层感知机之后，更深网络靠反向传播训练。\n\nConnectionism\n它是连接主义早期最有代表性的模型之一。", "relations": { "neural-network": { "label": "是…的起点", "note": "它启发了后来的神经网络发展。" }, "mlp": { "label": "发展成…", "note": "多层版本补上了单层表达短板。" }, "backpropagation": { "label": "为…铺路", "note": "更复杂网络后来靠它来训练。" }, "connectionism": { "label": "属于…路线", "note": "它代表用神经元建模智能的思路。" } } } } }, { "id": "permission-fatigue", "name": "Permission fatigue", "layer": "L6", "era": "2024", "publishedAt": "2026-05-28T15:58:23.418Z", "relations": [ { "to": "data-privacy" }, { "to": "computer-use" }, { "to": "agent" }, { "to": "ai-anxiety" } ], "track": "today", "i18n": { "en": { "fullName": "Permission fatigue", "factExplain": "The numbness users feel when endless permission pop-ups make them just tap Allow.", "humanExplain": "Permission fatigue is like a toddler asking, “Can I? Can I?” every two minutes. By lunch, you tap Yes just to save your sandwich.\n\nYou meet it in apps, plug-ins, and AI tools. You stop reading the boxes, and privacy risks slip by.", "humanExplainDisplay": "Permission fatigue is like a ==toddler asking, “Can I?”==\nevery two minutes.\nBy lunch, you tap ==Yes==\njust to save your sandwich.\n\nYou meet it in apps, plug-ins, and AI tools.\nYou stop reading the boxes,\nand privacy risks slip by.", "relationsNarrative": "Data-privacy\nPermission fatigue makes people less alert to privacy risks after too many prompts.\n\nComputer use\nComputer use tools may read the screen or control the mouse, so they ask often.\n\nAgent\nAn Agent may need email or calendar access, so its risks can be less obvious.\n\nAI-anxiety\nAI-anxiety grows when people do not understand permissions but keep getting interrupted.", "relations": { "data-privacy": { "label": "weakens … defenses", "note": "Too many Allow clicks can numb your privacy judgment." }, "computer-use": { "label": "often comes with …", "note": "Computer use tools often need more system permissions." }, "agent": { "label": "adds hidden risk to …", "note": "The more an Agent can do, the more sensitive access it may ask for." }, "ai-anxiety": { "label": "can worsen …", "note": "People feel uneasy when they keep clicking without understanding." } } }, "zh": { "fullName": "权限疲劳", "factExplain": "用户因频繁授权提示而变得麻木的现象。", "humanExplain": "权限疲劳就像外卖 App 连弹十个“允许吗”，打工人最后只想一路点同意。\n\n它会让人误批高风险操作，常见于智能体、插件和企业工具授权。", "humanExplainDisplay": "权限疲劳就像外卖 App ==连弹十个“允许吗”==，\n打工人最后只想==一路点同意==。\n\n它会让人误批高风险操作，\n常见于智能体、插件和企业工具授权。", "relationsNarrative": "Data-privacy\n权限疲劳会让用户在反复授权中变得麻木，降低对隐私风险的警觉。\n\nComputer use\ncomputer use 类工具常要读取屏幕、控制鼠标键盘，更容易把用户拖进频繁授权。\n\nAgent\nAgent 越能替你执行任务，往往越需要调用邮件、日历等权限，风险也更隐蔽。\n\nAI-anxiety\n当用户既不懂权限含义又被不断打扰时，容易对 AI 产品产生额外焦虑。", "relations": { "data-privacy": { "label": "削弱…防线", "note": "点多了授权，隐私判断会变麻木。" }, "computer-use": { "label": "常伴随…出现", "note": "代操作类工具往往需要更多系统权限。" }, "agent": { "label": "给…埋风险", "note": "越会替你干活，越常申请敏感权限。" }, "ai-anxiety": { "label": "会加重…", "note": "看不懂又点不停，最容易生出不安。" } } } } }, { "id": "perplexity", "name": "Perplexity", "layer": "L6", "era": "1970s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "rag" }, { "to": "llm" }, { "to": "information-retrieval" }, { "to": "hallucination" } ], "track": "history", "i18n": { "en": { "fullName": "AI Answer Search Engine", "factExplain": "A search tool that uses AI to write answers with sources.", "humanExplain": "Perplexity is like Google in a tiny detective hat. It gives you an answer, then points to its clues.\n\nUse it to research a topic or follow the news. The links help you check the sources.", "humanExplainDisplay": "Perplexity is like Google\nin a ==tiny detective hat==.\nIt gives you an answer,\nthen ==points to its clues==.\n\nUse it to research a topic\nor follow the news.\nThe links help you\ncheck the sources.", "relationsNarrative": "RAG\nPerplexity uses RAG to search pages before it answers.\n\nLLM\nAn LLM reads the sources and writes the answer.\n\nIR\nIR gives Perplexity its web sources and ranking base.\n\nHallucination\nSources can cut Hallucination, but they do not stop it forever.", "relations": { "rag": { "label": "uses … to answer", "note": "It searches the web first, then answers from those pages." }, "llm": { "label": "writes with …", "note": "The LLM turns source pages into a clear answer." }, "information-retrieval": { "label": "built on …", "note": "IR finds and ranks web pages for it." }, "hallucination": { "label": "reduces … with sources", "note": "Sources help, but it can still misread or make things up." } } }, "zh": { "fullName": "AI 答案搜索引擎", "factExplain": "一款用大模型生成带引用答案的搜索产品。", "humanExplain": "Perplexity像带路的导游：不让你自己翻地图找半天，直接带到答案，还指给你看依据。\n\n用于查资料、追新闻，带引用便于核对来源。", "humanExplainDisplay": "Perplexity像\n==带路的导游==：\n不让你自己翻地图，\n直接带到答案还==指给你看依据==。\n\n用于查资料、追新闻，\n带引用\n便于核对来源。", "relationsNarrative": "RAG\n它常先检索网页，再让模型依据资料回答。\n\nLLM\n大模型负责理解问题，并组织自然语言答案。\n\nInformation Retrieval\n传统检索提供网页来源和排序基础。\n\nHallucination\n引用能降低胡编，但不能保证永远正确。", "relations": { "rag": { "label": "用…生成答案", "note": "先找网页，再让模型按资料作答。" }, "llm": { "label": "调用…写答案", "note": "大模型把资料写成可读答案。" }, "information-retrieval": { "label": "建立在…之上", "note": "网页检索仍是它的信息入口。" }, "hallucination": { "label": "用引用降低…", "note": "有引用，也可能读错或编错。" } } } } }, { "id": "personal-ai-apps", "name": "Personal AI apps", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "chatgpt" }, { "to": "copilot" }, { "to": "api" }, { "to": "ai-device-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Personal AI Apps", "factExplain": "AI software made for one person’s everyday tasks.", "humanExplain": "A personal AI app is like an eager friend in your group chat. You ask for one sentence, and it brings a draft and bullet points.\n\nYou meet it in chat, notes, writing, and work apps. It helps you finish small tasks faster.", "humanExplainDisplay": "A personal AI app is like an ==eager friend==\nin your group chat.\nYou ask for one sentence,\nand it brings a ==draft and bullet points==.\n\nYou meet it in chat, notes,\nwriting, and work apps.\nIt helps you finish small tasks faster.", "relationsNarrative": "ChatGPT\nChatGPT is the most common chat-style doorway for personal AI apps.\n\nCopilot\nCopilot shows personal AI inside work and teamwork tools.\n\nAPI\nMany personal AI apps use APIs to connect to model power.\n\nAI device\nPersonal AI apps are also starting to run right on devices.", "relations": { "chatgpt": { "label": "often appears as …", "note": "ChatGPT is a common chat doorway into personal AI." }, "copilot": { "label": "spreads into …", "note": "Copilot shows how personal AI fits into work tools." }, "api": { "label": "often uses …", "note": "Many apps use APIs to add model power fast." }, "ai-device-ai": { "label": "can run on …", "note": "Personal AI apps are moving from the cloud onto devices." } } }, "zh": { "fullName": "个人 AI 应用", "factExplain": "面向个人日常任务的 AI 应用软件。", "humanExplain": "好比你书包里塞了个学霸同桌：题没念完，它已经把提纲、草稿和下半段答案都给你递过来了。\n\n常见于聊天、笔记、写作和办公，帮个人更快处理零碎任务。", "humanExplainDisplay": "好比你书包里塞了个==学霸同桌==：\n题没念完，\n它已经把提纲、草稿和\n下半段==答案都递过来==了。\n\n常见于聊天、\n笔记、写作和办公，\n帮个人更快处理零碎任务。", "relationsNarrative": "ChatGPT\n它是个人 AI 应用最典型的聊天式入口。\n\nCopilot\nCopilot 代表它在办公协作里的产品形态。\n\nAPI\n很多个人应用通过 API 接入模型能力。\n\nAI device\n这类应用也越来越多直接跑在终端设备上。", "relations": { "chatgpt": { "label": "典型形态是…", "note": "它常以聊天助手形式进入个人生活。" }, "copilot": { "label": "延伸到…场景", "note": "从通用助手走向办公协作入口。" }, "api": { "label": "常调用…能力", "note": "很多应用靠模型接口快速搭建。" }, "ai-device-ai": { "label": "也可跑在…上", "note": "个人应用正从云端延伸到终端设备。" } } } } }, { "id": "phrase-based-statistical-machine-translation", "name": "Phrase-Based SMT", "layer": "L4", "era": "2003", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "statistical-machine-translation" }, { "to": "machine-translation" }, { "to": "bleu" }, { "to": "beam-search" } ], "track": "history", "i18n": { "en": { "fullName": "Phrase-Based Statistical Machine Translation", "factExplain": "A statistical translation method that builds output from likely phrase chunks.", "humanExplain": "Phrase-Based SMT is fridge-magnet translation. It slides little word chunks around and picks the mix with the best score.\n\nIt powered many translation systems before neural AI took over. It often used Beam Search to find a high-scoring sentence.", "humanExplainDisplay": "Phrase-Based SMT is ==fridge-magnet translation==.\nIt slides ==little word chunks== around\nand picks the mix with the best score.\n\nIt powered many translation systems\nbefore neural AI took over.\nIt often used Beam Search\nto find a high-scoring sentence.", "relationsNarrative": "SMT\nPhrase-Based SMT was a main branch of SMT, built around phrase chunks.\n\nMT\nPhrase-Based SMT helped MT move from hand-made rules to data.\n\nBLEU\nBLEU was a common ruler for tuning it and ranking results.\n\nBeam Search\nBeam Search helped it find a high-scoring mix of phrase chunks.", "relations": { "statistical-machine-translation": { "label": "belongs to …", "note": "It was one of the main paths in the SMT era." }, "machine-translation": { "label": "serves …", "note": "Its goal is still to translate one language into another." }, "bleu": { "label": "is often scored by …", "note": "BLEU was often used to compare its translation quality." }, "beam-search": { "label": "decodes with …", "note": "There are too many phrase mixes, so it searches for a strong one." } } }, "zh": { "fullName": "基于短语的统计机器翻译", "factExplain": "用短语片段概率组合译文的统计翻译方法。", "humanExplain": "短语 SMT 像煎饼摊备料：葱花薄脆成把抓，按概率拼成一张译文。\n\n曾是机器翻译主力，常配 Beam Search 找高分译文。", "humanExplainDisplay": "短语 SMT 像\n==煎饼摊备料==：\n葱花薄脆成把抓，\n拼成==一张译文==。\n\n曾是机器翻译主力，\n常配 Beam Search 找高分译文。", "relationsNarrative": "SMT\n它是 SMT 最主流的一支，从词对齐扩展到短语块。\n\nMT\n它是 MT 从规则时代走向数据时代的关键方法。\n\nBLEU\nBLEU 曾是它调参和比赛的常用尺子。\n\nBeam Search\n它常用 Beam Search 在候选译文里找高分组合。", "relations": { "statistical-machine-translation": { "label": "属于…", "note": "它是 SMT 时代最主流的路线之一。" }, "machine-translation": { "label": "服务于…", "note": "目标仍是把一种语言自动译成另一种。" }, "bleu": { "label": "常用…评测", "note": "BLEU 常被用来比较译文质量。" }, "beam-search": { "label": "用…解码", "note": "短语组合太多，需要搜索高分译文。" } } } } }, { "id": "physical-ai-ai", "name": "Physical AI", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "multimodal" }, { "to": "world-model" }, { "to": "robotaxi" } ], "track": "today", "i18n": { "en": { "fullName": "Physical AI", "factExplain": "AI that can sense the real world and act in it.", "humanExplain": "A normal AI lives in a laptop. Physical AI has to cross the kitchen without stepping on the cat.\n\nYou meet it in robots and self-driving cars. It reads the world, then moves in it.", "humanExplainDisplay": "A normal AI lives in a ==laptop==.\nPhysical AI has to cross the ==kitchen==\nwithout stepping on the cat.\n\nYou meet it in robots and self-driving cars.\nIt reads the world,\nthen moves in it.", "relationsNarrative": "Embodied AI\nPhysical AI and Embodied AI both sense and act in the real world.\n\nMultimodal AI\nPhysical AI often uses Multimodal skills to read images, sound, and space.\n\nWorld model\nA World model helps Physical AI predict the world and its own actions.\n\nRobotaxi\nA Robotaxi is a common real-world use of Physical AI.", "relations": { "embodied-ai": { "label": "often overlaps with …", "note": "Both put AI inside the real world." }, "multimodal": { "label": "senses through …", "note": "It needs images, sound, and other real-world signals." }, "world-model": { "label": "models with …", "note": "It uses a world model to guess what may happen next." }, "robotaxi": { "label": "shows up as …", "note": "Self-driving cars are a clear use of Physical AI." } } }, "zh": { "fullName": "物理 AI", "factExplain": "能感知并作用于现实世界的 AI。", "humanExplain": "它不再只会网上冲浪，而是像新手学骑共享单车：得自己看路、拐弯、刹车，摔了真疼。\n\n多见于机器人和自动驾驶，重点是感知环境并实际行动。", "humanExplainDisplay": "它不再只会网上冲浪，\n而是像新手学骑==共享单车==：\n得自己看路、拐弯、刹车，\n摔了==真疼==。\n\n多见于机器人和自动驾驶，\n重点是感知环境并实际行动。", "relationsNarrative": "Embodied AI\n两者高度相关，都强调 AI 在现实环境中感知和行动。\n\nMultimodal AI\n它通常依赖多模态能力来理解图像、声音和空间信息。\n\nWorld model\n世界模型帮助它预测环境变化与行动后果。\n\nRobotaxi\n自动驾驶是物理 AI 的典型落地形态之一。", "relations": { "embodied-ai": { "label": "常与…重叠", "note": "两者都强调 AI 进入真实环境。" }, "multimodal": { "label": "依赖…感知", "note": "要看懂图像声音等现实信号。" }, "world-model": { "label": "常用…建模", "note": "需要预测环境变化与行动结果。" }, "robotaxi": { "label": "落地到…", "note": "自动驾驶是它的典型应用。" } } } } }, { "id": "physical-symbol-system-hypothesis", "name": "Physical Symbol System Hypothesis", "layer": "L1", "era": "1976", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "symbolic-ai" }, { "to": "production-system" }, { "to": "knowledge-representation" }, { "to": "connectionism" } ], "track": "history", "i18n": { "en": { "fullName": "Physical Symbol System Hypothesis", "factExplain": "A claim: general intelligence needs symbol rules, and symbol rules are enough.", "humanExplain": "It treats the mind like LEGO with a bossy manual. Every brick has a name, and the manual runs the show.\n\nYou meet it behind Symbolic AI and old expert systems. Connectionism pushes back with neural nets.", "humanExplainDisplay": "It treats the mind like ==LEGO==\nwith a ==bossy manual==.\nEvery brick has a name,\nand the manual runs the show.\n\nYou meet it behind Symbolic AI\nand old expert systems.\nConnectionism pushes back\nwith neural nets.", "relationsNarrative": "Symbolic AI\nThis hypothesis gave Symbolic AI its main theory of intelligence.\n\nProduction\nProduction systems are often used to run symbol rules.\n\nKR\nKR turns the world into symbols a system can use.\n\nConnectionism\nConnectionism challenges the pure symbol view with neural networks.", "relations": { "symbolic-ai": { "label": "grounds …", "note": "It gave Symbolic AI its main theory of intelligence." }, "production-system": { "label": "runs rules with …", "note": "Production systems make symbol rules actually run." }, "knowledge-representation": { "label": "codes the world with …", "note": "KR turns the world into symbols a system can use." }, "connectionism": { "label": "is challenged by …", "note": "Connectionism argues intelligence can come from neural networks, not just symbols." } } }, "zh": { "fullName": "物理符号系统假说", "factExplain": "认为符号操作是通用智能的充分必要条件。", "humanExplain": "它信智能像下棋背棋谱：棋子格子全编号，照规则推演就像会思考。\n\n它支撑符号 AI、专家系统，也被神经网络路线持续挑战。", "humanExplainDisplay": "它信智能像\n==下棋背棋谱==：\n棋子格子全编号，\n照规则推演就像会思考。\n\n它支撑符号 AI、专家系统，\n也被神经网络路线\n持续挑战。", "relationsNarrative": "Symbolic AI\n这个假说为符号主义智能观提供理论底座。\n\nProduction\n产生式系统常被用来实现符号规则操作。\n\nKnowledge Representation\n知识表示把世界整理成可操作的符号。\n\nConnectionism\n连接主义用神经网络挑战纯符号路线。", "relations": { "symbolic-ai": { "label": "奠基…", "note": "它给符号主义提供理论底座。" }, "production-system": { "label": "用…落地规则", "note": "产生式系统把符号规则跑起来。" }, "knowledge-representation": { "label": "依赖…编码世界", "note": "知识表示把世界变成符号。" }, "connectionism": { "label": "被…挑战", "note": "神经网络路线反驳纯符号观。" } } } } }, { "id": "planning-graph", "name": "Planning Graph", "layer": "L2", "era": "1995", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "automated-planning" }, { "to": "strips" }, { "to": "heuristic-search" }, { "to": "pddl" } ], "track": "history", "seo": { "zh": { "title": "规划图 是什么?围棋算路棋谱,一文看懂 — AI Rookies", "description": "一种分层表示动作与状态可达性的规划结构。人话版一分钟讲明白。" }, "en": { "title": "What Is Planning Graph? Robot Maid Chore Ladder", "description": "A layered planning map of possible actions, states, and conflicts. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Planning Graph", "factExplain": "A layered planning map of possible actions, states, and conflicts.", "humanExplain": "A Planning Graph is like a chore ladder for a robot maid. Each rung shows possible moves, plus bad combos like “mop” and “muddy dog.”\n\nIt helps a planner find a working action list faster. You meet it in Planning and Heuristic Search.", "humanExplainDisplay": "A Planning Graph is like a ==chore ladder== for a robot maid.\nEach rung shows possible moves,\nplus bad combos like ==“mop” and “muddy dog.”==\n\nIt helps a planner find\na working action list faster.\nYou meet it in Planning\nand Heuristic Search.", "relationsNarrative": "Planning\nA Planning Graph lays out actions and states for Planning.\n\nSTRIPS\nA Planning Graph often uses STRIPS needs and effects to build layers.\n\nHeuristic Search\nA relaxed Planning Graph estimates distance for Heuristic Search.\n\nPDDL\nPDDL often describes the task a Planning Graph must handle.", "relations": { "automated-planning": { "label": "supports …", "note": "Planning graphs are core tools in classic planning." }, "strips": { "label": "uses … action rules", "note": "STRIPS says what an action needs and what it changes." }, "heuristic-search": { "label": "guides …", "note": "Relaxed planning graphs help guess the distance to the goal." }, "pddl": { "label": "uses tasks from …", "note": "PDDL often describes the planning task to solve." } } }, "zh": { "fullName": "规划图", "factExplain": "一种分层表示动作与状态可达性的规划结构。", "humanExplain": "规划图像下围棋算路：哪些棋能连，哪些会互堵，先摊成一层层棋谱。\n\n用于自动规划和启发式搜索，更快找到可行动作序列。", "humanExplainDisplay": "规划图像下围棋算路：\n哪些棋==能连==，\n哪些会==互堵==，\n先摊成一层层棋谱。\n\n用于自动规划和启发式搜索，\n更快找到可行动作序列。", "relationsNarrative": "Planning\n规划图是自动规划里展开动作与状态的结构。\n\nSTRIPS\n规划图常用 STRIPS 的前提和效果来建层。\n\nHeuristic Search\n松弛规划图可为启发式搜索估算目标距离。\n\nPDDL\nPDDL 常描述规划图要处理的规划任务。", "relations": { "automated-planning": { "label": "支撑…求解", "note": "规划图是经典规划的核心结构。" }, "strips": { "label": "承接…动作模型", "note": "动作前提和效果常按 STRIPS 表示。" }, "heuristic-search": { "label": "给…提供估价", "note": "松弛规划图常用来估算距离。" }, "pddl": { "label": "读取…任务描述", "note": "PDDL 常描述规划图要解的任务。" } } } } }, { "id": "policy-gradient", "name": "Policy Gradient", "layer": "L2", "era": "1992", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "rlhf" }, { "to": "sgd" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Policy Gradient", "factExplain": "A reinforcement learning method that improves choices to earn higher long-term reward.", "humanExplain": "Policy Gradient coaches AI like a Mario Kart parent. It skips every tiny lecture and rewards the whole good lap.\n\nIt updates the AI’s action rules, so rewarded choices happen more often. You meet it in robot control and RLHF.", "humanExplainDisplay": "Policy Gradient coaches AI like a ==Mario Kart parent==.\nIt skips every tiny lecture\nand rewards the ==whole good lap==.\n\nIt updates the AI’s action rules,\nso rewarded choices happen more often.\nYou meet it in robot control and RLHF.", "relationsNarrative": "RLHF\nRLHF often uses Policy Gradient to move the model toward rewarded answers.\n\nSGD\nPolicy Gradient finds the gradient, then SGD often updates the weights.\n\nAgent\nPolicy Gradient trains an Agent through many tries and rewards.", "relations": { "rlhf": { "label": "often trains …", "note": "RLHF often uses Policy Gradient to update a policy from rewards." }, "sgd": { "label": "updates with …", "note": "SGD usually uses its gradient to update the weights." }, "agent": { "label": "trains … decisions", "note": "It suits Agents that make many choices in a row." } } }, "zh": { "fullName": "策略梯度", "factExplain": "一种直接优化策略、提升长期回报的强化学习方法。", "humanExplain": "它教 AI 踢球不盯每脚姿势，只看这回合最后进没进：进了，就把刚才那套跑位和出脚多练几遍。\n\n它常用于机器人控制和 RLHF，让模型根据反馈逐步学会更好的决策。", "humanExplainDisplay": "它教 AI 踢球\n不盯每脚姿势，\n只看这回合最后==进没进==：\n进了，就把刚才那套\n==跑位和出脚==多练几遍。\n\n它常用于机器人控制和 RLHF，\n让模型根据反馈逐步学会更好的决策。", "relationsNarrative": "RLHF\nRLHF 常用它根据奖励信号，调整模型偏好的方向。\n\nSGD\n它算出梯度后，通常还是靠梯度下降更新参数。\n\nAgent\n它适合训练需要连续决策的智能体不断试错。", "relations": { "rlhf": { "label": "常用于…训练", "note": "RLHF 常用它根据奖励更新策略。" }, "sgd": { "label": "靠…做更新", "note": "它通常仍用梯度下降来改参数。" }, "agent": { "label": "可训练…决策", "note": "适合连续行动与试错式任务。" } } } } }, { "id": "policy-iteration", "name": "Policy Iteration", "layer": "L2", "era": "1957", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "markov-decision-process" }, { "to": "bellman-equation" }, { "to": "value-iteration" } ], "track": "history", "i18n": { "en": { "fullName": "Policy Iteration", "factExplain": "An RL method that scores a policy, improves it, then repeats.", "humanExplain": "Policy Iteration is a basketball coach at practice. First run the play. Then fix the ugly part.\n\nIn AI, it scores the current policy. Then it improves the policy and repeats. It is common in RL planning with a known model.", "humanExplainDisplay": "Policy Iteration is a ==basketball coach== at practice.\nFirst run the play.\nThen ==fix the ugly part==.\n\nIn AI, it scores the current policy.\nThen it improves the policy and repeats.\nIt is common in RL planning\nwith a known model.", "relationsNarrative": "RL\nPolicy Iteration is a classic RL method for finding better decisions.\n\nMDP\nPolicy Iteration usually uses an MDP to define states, actions, and rewards.\n\nBellman Equation\nPolicy evaluation uses the Bellman Eq to compute values.\n\nValue Iteration\nPolicy Iteration has the same goal as Value Iteration, but it separates scoring and improving.", "relations": { "reinforcement-learning": { "label": "is a classic … method", "note": "Policy Iteration is a basic dynamic programming method in RL." }, "markov-decision-process": { "label": "solves … policies", "note": "It often repeats policy scoring and improvement inside an MDP." }, "bellman-equation": { "label": "scores with …", "note": "Policy evaluation uses the Bellman Eq to compute values." }, "value-iteration": { "label": "contrasts with …", "note": "Both seek the best policy, but their update rhythm differs." } } }, "zh": { "fullName": "策略迭代", "factExplain": "一种交替做策略评估与改进的强化学习求解法。", "humanExplain": "像武侠门派拆招：先验这套招法能不能赢，再把破绽改掉，下一轮换更强的出手。\n\n常用于已知环境模型的规划，靠反复评估和改进逼近更优决策。", "humanExplainDisplay": "像武侠门派拆招：\n先验这套招法==能不能赢==，\n再把破绽改掉，\n下一轮换==更强的出手==。\n\n常用于已知环境模型的规划，\n靠反复评估和改进逼近更优决策。", "relationsNarrative": "Reinforcement-learning\n它是强化学习里求最优决策的经典方法。\n\nMarkov-decision-process\n它通常在 MDP 框架下定义状态、动作和回报。\n\nBellman-equation\n策略评估这一步，要靠贝尔曼方程计算价值。\n\nValue-iteration\n它和价值迭代目标相同，但分开评估与改进。", "relations": { "reinforcement-learning": { "label": "属于…经典方法", "note": "它是强化学习里的基础动态规划法。" }, "markov-decision-process": { "label": "求解…最优策略", "note": "通常在 MDP 设定下反复评估与改进。" }, "bellman-equation": { "label": "用…做评估", "note": "策略评估要靠贝尔曼方程算价值。" }, "value-iteration": { "label": "与…相对照", "note": "两者都求最优策略，但更新节奏不同。" } } } } }, { "id": "positional-encoding", "name": "Positional Encoding", "layer": "L3", "era": "2017", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "self-attention" }, { "to": "embedding" } ], "track": "history", "i18n": { "en": { "fullName": "Positional Encoding", "factExplain": "A way to give each sequence item its position.", "humanExplain": "Without positional encoding, a sentence is fridge magnets in a junk drawer. All the words are there, but 'dog bites man' can turn into 'man bites dog.'\n\nIt gives each word a seat number. Then models keep order in long writing and code.", "humanExplainDisplay": "Without positional encoding,\na sentence is ==fridge magnets==\nin a junk drawer.\nAll the words are there,\nbut ==dog bites man==\ncan turn into 'man bites dog.'\n\nIt gives each word a seat number.\nThen models keep order\nin long writing and code.", "relationsNarrative": "Transformer\nPositional Encoding lets a Transformer know each word's place.\n\nSelf-Attention\nSelf-Attention can compare words, but it needs help with order.\n\nEmbedding\nPositional Encoding is often added to the word Embedding.", "relations": { "transformer": { "label": "lets … sense order", "note": "Transformer uses it to know each word's place." }, "self-attention": { "label": "adds order to …", "note": "Self-Attention does not know before and after by itself." }, "embedding": { "label": "adds onto …", "note": "The position vector is often added to the word vector." } } }, "zh": { "fullName": "位置编码", "factExplain": "一种为序列元素加入位置信息的编码方式。", "humanExplain": "没有位置编码，句子就像早高峰地铁：词全挤上车，乱成一锅粥，没人知道谁先谁后。\n\n它补上顺序感，让模型读语序、长文本和代码不串行。", "humanExplainDisplay": "没有位置编码，\n句子就像==早高峰地铁==：\n词全挤上车，乱成一锅粥，\n没人知道==谁先谁后==。\n\n它补上顺序感，\n让模型读语序、长文本和代码\n不串行。", "relationsNarrative": "Transformer\n位置编码让 Transformer 知道词在序列中的位置。\n\nSelf-Attention\n自注意力看全局关系，但本身不带先后顺序。\n\nEmbedding\n它常与词向量相加，一起送进模型。", "relations": { "transformer": { "label": "让…感知顺序", "note": "Transformer 借它补上序列位置信息。" }, "self-attention": { "label": "补足…的顺序感", "note": "自注意力本身不天然区分前后。" }, "embedding": { "label": "叠加到…上", "note": "位置向量常与词向量相加输入模型。" } } } } }, { "id": "post-training", "name": "Post-training", "layer": "L2", "era": "2020s", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "fine-tuning" }, { "to": "instruction-tuning" }, { "to": "rlhf" } ], "track": "today", "i18n": { "en": { "fullName": "Post-training", "factExplain": "The stage after pretraining that shapes a model’s behavior and preferences.", "humanExplain": "Post-training is finishing school for an AI. It already read the library. Now it learns not to interrupt Grandma.\n\nIt tunes safety and preferences before an assistant goes live. You feel it when a chatbot follows instructions and stays polite.", "humanExplainDisplay": "Post-training is ==finishing school== for an AI.\nIt already read the library.\nNow it learns ==not to interrupt Grandma==.\n\nIt tunes safety and preferences\nbefore an assistant goes live.\nYou feel it when a chatbot\nfollows instructions and stays polite.", "relationsNarrative": "Pretraining\nPretraining teaches broad knowledge first. Post-training then shapes behavior and preferences.\n\nFine-tuning\nFine-tuning is one common tool used during post-training.\n\nInstruction Tuning\nInstruction Tuning uses examples to teach the model to follow requests.\n\nRLHF\nRLHF uses human feedback to make answers fit human preferences.", "relations": { "pretraining": { "label": "comes after …", "note": "First the model learns broad knowledge. Then post-training shapes behavior." }, "fine-tuning": { "label": "often includes …", "note": "Fine-tuning is a common tool in post-training." }, "instruction-tuning": { "label": "uses … to follow requests", "note": "Instruction Tuning teaches the model to follow requests." }, "rlhf": { "label": "uses … to match preferences", "note": "RLHF uses human feedback to shape better replies." } } }, "zh": { "fullName": "后训练", "factExplain": "预训练后调校模型行为与偏好的阶段。", "humanExplain": "后训练就是相亲前的礼仪课：本事已有，再教模型别抢话、会接梗。\n\n用于调安全和偏好，常在助手上线前完成。", "humanExplainDisplay": "后训练就是\n==相亲前的礼仪课==：\n本事已有，\n再教模型==别抢话、会接梗==。\n\n用于调安全和偏好，\n常在助手上线前完成。", "relationsNarrative": "Pretraining\n预训练先学通识，后训练再调行为和偏好。\n\nFine-tuning\n微调是后训练里最常见的技术手段之一。\n\nInstruction Tuning\n指令微调用样例教模型更好地听人话。\n\nRLHF\nRLHF 用人类反馈把回答调得更符合偏好。", "relations": { "pretraining": { "label": "接在…之后", "note": "先学通识，再做行为调校。" }, "fine-tuning": { "label": "包含…环节", "note": "微调常是后训练的基本手艺。" }, "instruction-tuning": { "label": "用…教听指令", "note": "指令微调让模型更会按要求答。" }, "rlhf": { "label": "借…对齐偏好", "note": "人类反馈常用于调出更合适回答。" } } } } }, { "id": "prefill", "name": "Prefill", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "kv-cache" }, { "to": "llm-inference-engine" }, { "to": "context-window" } ], "track": "today", "i18n": { "en": { "fullName": "Prefill", "factExplain": "The step where an AI reads the input and builds a KV cache.", "humanExplain": "Prefill is a pizza shop reading your giant order first. No slice is out yet, but the sticky notes are ready.\n\nYou meet it in chatbots and long document Q&A. It shapes the wait for the first word, and part of the cost.", "humanExplainDisplay": "Prefill is a ==pizza shop==\nreading your giant order first.\nNo slice is out yet,\nbut the ==sticky notes are ready==.\n\nYou meet it in chatbots\nand long document Q&A.\nIt shapes the wait for the first word,\nand part of the cost.", "relationsNarrative": "KV cache\nPrefill turns the whole prompt into a KV cache for later words.\n\nInference engine\nThe inference engine schedules prefill and word-by-word generation.\n\nContext-window\nA longer context-window gives prefill more text to read.", "relations": { "kv-cache": { "label": "builds …", "note": "Prefill turns the input prompt into the KV cache." }, "llm-inference-engine": { "label": "is scheduled by …", "note": "The engine schedules prefill and word-by-word generation." }, "context-window": { "label": "is affected by …", "note": "A longer context makes prefill process more text." } } }, "zh": { "fullName": "预填充", "factExplain": "推理时先处理输入并生成 KV 缓存的阶段。", "humanExplain": "预填充像答题前先通读全文：一个字还没写，重点已经全画好线。\n\n用于聊天和长文问答，决定首字等待，也影响部分成本。", "humanExplainDisplay": "预填充像答题前\n==先通读全文==：\n一个字还没写，\n重点已经==全画好线==。\n\n用于聊天和长文问答，\n决定首字等待，\n也影响部分成本。", "relationsNarrative": "KV cache\n预填充把整段提示算成 KV 缓存，供后续生成复用。\n\nInference engine\n推理引擎负责调度预填充和逐词生成两阶段。\n\nContext-window\n上下文越长，预填充要处理的内容越多。", "relations": { "kv-cache": { "label": "生成…", "note": "预填充把输入提示算成缓存。" }, "llm-inference-engine": { "label": "由…调度", "note": "引擎安排预填充与逐词生成。" }, "context-window": { "label": "受…影响", "note": "上下文越长，预填充越慢。" } } } } }, { "id": "pretraining", "name": "Pretraining", "layer": "L2", "era": "2018", "publishedAt": "2026-05-23T08:35:00Z", "relations": [ { "to": "foundation-model" }, { "to": "parameter" }, { "to": "scaling-law" }, { "to": "gpu" } ], "i18n": { "en": { "fullName": "Pretraining", "factExplain": "A big early training stage gives a model general skills from huge data.", "humanExplain": "Pretraining is like a kid binge-watching every how-to video online. No chores yet. First, fill the brain.\n\nIt helps the model learn language, plus facts and patterns. Fine-tuning and real apps build on this base.", "humanExplainDisplay": "Pretraining is like a kid\n==binge-watching every how-to video== online.\n==No chores yet==.\nFirst, fill the brain.\n\nIt helps the model learn language,\nplus facts and patterns.\nFine-tuning and real apps\nbuild on this base.", "relationsNarrative": "Foundation-model\nPretraining gives a Foundation-model most of its general skills.\n\nParameter\nPretraining keeps updating Parameters with huge amounts of data.\n\nScaling-law\nScaling-law pushes pretraining toward more data and bigger runs.\n\nGPU\nPretraining needs more GPU power as it gets bigger.", "relations": { "foundation-model": { "label": "gives … general skills", "note": "A Foundation-model gets most general skills from large-scale pretraining." }, "parameter": { "label": "updates many …", "note": "Pretraining changes many Parameters as the model learns from huge data." }, "scaling-law": { "label": "follows …", "note": "Scaling-law helps explain why bigger pretraining often improves models." }, "gpu": { "label": "depends on …", "note": "Bigger pretraining needs more GPU power." } } }, "zh": { "fullName": "预训练", "factExplain": "用大规模数据训练模型获得通用能力的阶段。", "humanExplain": "预训练像先把孩子扔进巨型图书馆，能不能成才另说，见识先堆满。\n\n它让模型学会语言、知识和模式，是后续微调和应用的地基。", "humanExplainDisplay": "预训练像先把模型\n扔进==巨型图书馆==。\n\n能不能成才另说，\n见识先堆满。\n\n它让模型学会语言、\n知识和模式。\n后面的微调和应用，\n都站在这块地基上。", "relationsNarrative": "Foundation-model\nFoundation-model 的通用能力主要来自大规模 Pretraining。\n\nParameter\nPretraining 通过海量数据不断更新 Parameter。\n\nScaling-law\nScaling-law 推动了持续扩大的 Pretraining 规模与数据投入。\n\nGPU\nPretraining 规模越大，对 GPU 算力的需求越高。", "relations": { "foundation-model": { "label": "让…获得通用能力" }, "parameter": { "label": "大量调整…" }, "scaling-law": { "label": "遵循…" }, "gpu": { "label": "依赖…" } } } } }, { "id": "principal-component-analysis", "name": "PCA", "layer": "L2", "era": "1950", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "dimensionality-reduction" }, { "to": "feature-selection" }, { "to": "unsupervised-learning" }, { "to": "autoencoder" } ], "track": "history", "i18n": { "en": { "fullName": "Principal Component Analysis", "factExplain": "PCA shrinks data into a few new features with the biggest patterns.", "humanExplain": "PCA is like squinting at a messy school photo. You miss the braces, but you still spot the class clown.\n\nIt turns many columns into a few new ones for simple charts or noise cleanup. Squeeze too hard, and small details vanish.", "humanExplainDisplay": "PCA is like ==squinting== at a messy school photo.\nYou miss the braces,\nbut you still spot the ==class clown==.\n\nIt turns many columns into a few new ones\nfor simple charts or noise cleanup.\nSqueeze too hard,\nand small details vanish.", "relationsNarrative": "Dim. Reduction\nPCA uses a few main parts to sum up the data.\n\nFeature Selection\nPCA mixes old features into new ones instead of picking them.\n\nUnsupervised Learning\nPCA needs no labels and finds the biggest changes in the data.\n\nAutoencoder\nPCA and an Autoencoder both compress data, but an Autoencoder is often more flexible.", "relations": { "dimensionality-reduction": { "label": "is a classic … method", "note": "PCA is one of the classic ways to reduce dimensions." }, "feature-selection": { "label": "differs from …", "note": "PCA makes new features instead of picking old ones." }, "unsupervised-learning": { "label": "often counts as …", "note": "PCA uses no labels and looks at the data shape." }, "autoencoder": { "label": "can be compared with …", "note": "Both compress data, but PCA is straight-line and Autoencoders are more flexible." } } }, "zh": { "fullName": "主成分分析", "factExplain": "把数据压缩到少数最主要变化方向的降维方法。", "humanExplain": "老中医看舌脉，不会把几十个细枝末节平铺开讲，先抓住最主的几路症候。\n\n它常用于降维、可视化和去噪，但压太狠会丢细节。", "humanExplainDisplay": "老中医看舌脉，\n不会把几十个细枝末节平铺开讲，\n先抓住最主的\n==几路症候==。\n\n它常用于降维、可视化\n和去噪，\n但压太狠会丢细节。", "relationsNarrative": "Dimensionality Reduction\nPCA 是降维里的经典方法，用少数主成分概括数据。\n\nFeature Selection\n它不是挑原特征保留，而是把原特征重新组合。\n\nUnsupervised Learning\n它不需要标签，主要从数据本身找变化最大的方向。\n\nAutoencoder\n两者都能做压缩表示，但一个线性、一个通常更灵活。", "relations": { "dimensionality-reduction": { "label": "属于…方法", "note": "它是最经典的降维方法之一。" }, "feature-selection": { "label": "不同于…", "note": "它造新特征，不是挑旧特征。" }, "unsupervised-learning": { "label": "常归入…", "note": "它不靠标签，只看数据结构。" }, "autoencoder": { "label": "可对比…", "note": "两者都能压缩表示，但路线不同。" } } } } }, { "id": "probabilistic-context-free-grammar", "name": "PCFG", "layer": "L2", "era": "1969", "publishedAt": "2026-06-30T04:00:00.000Z", "relations": [ { "to": "syntactic-parsing" }, { "to": "probabilistic-graphical-model" }, { "to": "penn-treebank" } ], "track": "history", "i18n": { "en": { "fullName": "Probabilistic Context-Free Grammar", "factExplain": "A grammar model that puts probabilities on grammar rules.", "humanExplain": "PCFG is like a picky English teacher grading a messy sentence. It gives each possible reading a “yeah, probably” score.\n\nIt helps computers parse sentences and hear speech. It picks the most likely sentence structure.", "humanExplainDisplay": "PCFG is like a ==picky English teacher==\ngrading a messy sentence.\nIt gives each possible reading\na ==“yeah, probably” score==.\n\nIt helps computers parse sentences\nand hear speech.\nIt picks the most likely sentence structure.", "relationsNarrative": "Syntax Parse\nPCFG scores different syntax trees and helps choose the most likely one.\n\nPGM\nA PGM view can explain how PCFG generates structure with uncertainty.\n\nPenn Treebank\nPenn Treebank trees are often used to estimate PCFG rule probabilities.", "relations": { "syntactic-parsing": { "label": "scores …", "note": "PCFG gives each possible syntax tree a probability." }, "probabilistic-graphical-model": { "label": "shows uncertainty with …", "note": "PCFG uses probability to describe how structure is made." }, "penn-treebank": { "label": "learns rule odds from …", "note": "Treebank counts can estimate rule weights." } } }, "zh": { "fullName": "概率上下文无关文法", "factExplain": "给上下文无关文法规则分配概率的语法模型。", "humanExplain": "PCFG 像阅卷老师拆病句：不只判能不能通，还给每种读法估个靠谱分。\n\n用于句法分析、语音识别，在多种解析中挑最可能的。", "humanExplainDisplay": "PCFG 像==阅卷老师==拆病句：\n不只判能不能通，\n还给每种读法\n估个==靠谱分==。\n\n用于句法分析、\n语音识别，\n在多种解析中挑最可能的。", "relationsNarrative": "Syntax Parse\nPCFG 给不同句法树打分，帮解析器选最可能结构。\n\nPGM\nPCFG 可用概率图模型视角理解其生成过程。\n\nPenn Treebank\n树库里的标注句法树常用来估计规则概率。", "relations": { "syntactic-parsing": { "label": "为…打分", "note": "PCFG 给候选句法树分配概率。" }, "probabilistic-graphical-model": { "label": "借…表达不确定性", "note": "它用概率描述结构如何生成。" }, "penn-treebank": { "label": "从…学习规则概率", "note": "树库统计可估计规则权重。" } } } } }, { "id": "probabilistic-graphical-model", "name": "PGM", "layer": "L3", "era": "1980s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "bayesian-network" }, { "to": "belief-propagation" }, { "to": "variational-inference" }, { "to": "structural-causal-model" } ], "track": "history", "i18n": { "en": { "fullName": "Probabilistic Graphical Model", "factExplain": "A model that maps uncertain things and their probability links as a graph.", "humanExplain": "A PGM is a detective corkboard for messy guesses. The red string shows who affects whom, so nobody blames the dog for the Wi-Fi.\n\nYou meet it in diagnosis and forecasts. It can fill in missing facts and update the odds as clues arrive.", "humanExplainDisplay": "A PGM is a ==detective corkboard==\nfor messy guesses.\nThe ==red string== shows who affects whom,\nso nobody blames the dog for the Wi-Fi.\n\nYou meet it in diagnosis and forecasts.\nIt can fill in missing facts\nand update the odds as clues arrive.", "relationsNarrative": "Bayesian Network\nA Bayesian Network is one classic branch of PGM.\n\nBP\nBP updates beliefs by passing messages along the graph.\n\nVI\nVI gives an approximate answer when the graph is too complex.\n\nSCM\nSCM uses graph ideas, but it cares more about causes and interventions.", "relations": { "bayesian-network": { "label": "includes …", "note": "A Bayesian Network is a classic type of PGM." }, "belief-propagation": { "label": "infers with …", "note": "BP passes messages along the graph to update beliefs." }, "variational-inference": { "label": "approximates with …", "note": "VI helps when the graph is too hard to solve exactly." }, "structural-causal-model": { "label": "helps ground …", "note": "SCM borrows the graph idea but focuses on causes and actions." } } }, "zh": { "fullName": "Probabilistic Graphical Model／概率图模型", "factExplain": "用图结构表示随机变量及其概率依赖关系的模型。", "humanExplain": "它像小区群里传消息画线头：谁听谁的、谁会影响谁，顺着线往回捋，猜事就不靠瞎蒙。\n\n常用于诊断、预测和缺失信息推理；能边看证据边更新判断。", "humanExplainDisplay": "它像小区群里传消息\n画==线头==：\n谁听谁的、谁会影响谁，\n顺着线往回捋，\n猜事就不靠==瞎蒙==。\n\n常用于诊断、预测和\n缺失信息推理；\n能边看证据边更新判断。", "relationsNarrative": "Bayesian Network\n贝叶斯网络是概率图模型里最经典的一支。\n\nBelief-propagation\n它常用信念传播在图上更新各节点判断。\n\nVariational Inference\n图模型过复杂时，常用变分推断做近似求解。\n\nStructural Causal Model\n因果模型继承了图表示，但更强调干预与因果方向。", "relations": { "bayesian-network": { "label": "包含…这类模型", "note": "贝叶斯网络是其经典分支。" }, "belief-propagation": { "label": "用…做推断", "note": "图上传消息是常见求解办法。" }, "variational-inference": { "label": "常配合…近似", "note": "图太复杂时常靠近似推断。" }, "structural-causal-model": { "label": "为…提供基础", "note": "很多因果图建模借鉴其图表示。" } } } } }, { "id": "production-system", "name": "Production", "layer": "L3", "era": "1970s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "logic-programming" } ], "track": "history", "i18n": { "en": { "fullName": "Production System", "factExplain": "A way to reason using if-this-then-do-that rules.", "humanExplain": "A Production System is like a phone menu. Press 1 for support, press 2 for billing. Each press runs its own set path.\n\nIt uses clear if-this-then-do-that rules. It helps with automatic decisions and rule-based reasoning.", "humanExplainDisplay": "A Production System is like a ==phone menu==.\nPress 1 for support, press 2 for billing.\nEach press ==runs its own set path==.\n\nIt uses clear if-this-then-do-that rules.\nIt helps with automatic decisions\nand rule-based reasoning.", "relationsNarrative": "KR\nA Production System turns knowledge into matchable if-then rules.\n\nLogic\nBoth use symbolic reasoning, but they write rules in different ways.", "relations": { "knowledge-representation": { "label": "turns … into rules", "note": "It writes knowledge as rules the system can run." }, "logic-programming": { "label": "shares rules with …", "note": "Both use clear rules and symbolic reasoning." } } }, "zh": { "fullName": "Production System（产生式系统）", "factExplain": "用“条件-动作”规则推理与决策的方法。", "humanExplain": "像打客服的按键菜单：按 1 转人工、按 2 查账单，你一按对应键，它就自动走那条早写好的流程。\n\n适合规则明确的判断流程，常用来做自动决策和推理。", "humanExplainDisplay": "像打客服的==按键菜单==：\n按 1 转人工、按 2 查账单，\n你一按对应键，\n它就==自动走==那条早写好的流程。\n\n适合规则明确的判断流程，\n常用来做自动决策\n和推理。", "relationsNarrative": "Knowledge-representation\n它把知识表示成可匹配、可执行的条件动作规则。\n\nLogic-programming\n两者都走符号推理路线，但表达形式不同。", "relations": { "knowledge-representation": { "label": "实现…规则", "note": "它把知识写成可执行规则。" }, "logic-programming": { "label": "同属符号路线", "note": "两者都强调规则与显式推理。" } } } } }, { "id": "prolog", "name": "Prolog", "layer": "L5", "sublayer": "product", "era": "1972", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "logic-programming" }, { "to": "unification" }, { "to": "resolution-principle" }, { "to": "expert-system" } ], "track": "history", "i18n": { "en": { "fullName": "Prolog", "factExplain": "A programming language that reasons from facts and rules.", "humanExplain": "Prolog is like a detective with a corkboard. Give it clues and rules, and it follows the string.\n\nYou meet it in expert systems and proof tools. It turns “we know” into “so this follows.”", "humanExplainDisplay": "Prolog is like a ==detective with a corkboard==.\nGive it ==clues and rules==,\nand it follows the string.\n\nYou meet it in expert systems and proof tools.\nIt turns “we know” into\n“so this follows.”", "relationsNarrative": "Logic\nProlog is one of the classic languages for Logic.\n\nUnification\nUnification matches query variables with rules automatically.\n\nResolution\nResolution gives Prolog a base for automatic reasoning.\n\nExpert System\nExpert systems use this kind of rule reasoning to store knowledge.", "relations": { "logic-programming": { "label": "implements …", "note": "Prolog is a classic language for Logic." }, "unification": { "label": "matches variables with …", "note": "Unification lines up a query with the right rule." }, "resolution-principle": { "label": "reasons through …", "note": "Resolution is a key base for Prolog's automatic reasoning." }, "expert-system": { "label": "is used in …", "note": "Prolog fits expert systems with many rules." } } }, "zh": { "fullName": "逻辑编程语言", "factExplain": "一种用事实和规则进行逻辑推理的编程语言。", "humanExplain": "Prolog 像侦探玩剧本杀：给线索和规则，它自己顺藤摸瓜找凶手。\n\n常用于专家系统、定理证明，擅长把“已知”推成“所以”。", "humanExplainDisplay": "Prolog 像侦探玩\n==剧本杀==：\n给线索和规则，\n它自己==顺藤摸瓜==找凶手。\n\n常用于专家系统、定理证明，\n擅长把“已知”\n推成“所以”。", "relationsNarrative": "Logic Programming\nProlog 是逻辑编程最经典的代表语言之一。\n\nUnification\n合一让查询里的变量和规则自动对上。\n\nResolution\n归结为它的自动推理提供逻辑基础。\n\nExpert System\n专家系统常用这类规则推理表达知识。", "relations": { "logic-programming": { "label": "实现…范式", "note": "Prolog 是逻辑编程的代表语言。" }, "unification": { "label": "依靠…匹配变量", "note": "合一让查询和规则自动对上。" }, "resolution-principle": { "label": "借…推导答案", "note": "归结是其自动推理的重要基础。" }, "expert-system": { "label": "常用于…", "note": "它适合写规则密集的专家系统。" } } } } }, { "id": "prompt-engineering", "name": "Prompt-engineering", "layer": "L5", "sublayer": "architecture", "era": "2022", "publishedAt": "2026-05-23T10:35:00Z", "relations": [ { "to": "prompt" }, { "to": "llm" }, { "to": "chain-of-thought" }, { "to": "agent" } ], "i18n": { "en": { "fullName": "Prompt Engineering", "factExplain": "A way to write prompts so AI gives the result you want more often.", "humanExplain": "Prompt engineering is turning “just do it” into a clear recipe card. The AI stops guessing and stops serving soup with a fork.\n\nIt makes AI answer the same way more often, with neat formats. You use it in chatbots and Agents.", "humanExplainDisplay": "Prompt engineering is turning “just do it”\ninto a ==clear recipe card==.\nThe AI stops guessing\nand stops serving ==soup with a fork==.\n\nIt makes AI answer the same way more often,\nwith neat formats.\nYou use it in chatbots and Agents.", "relationsNarrative": "Prompt\nPrompt engineering uses clear structure to improve a Prompt.\n\nLLM\nPrompt engineering helps an LLM answer in a steadier way.\n\nChain-of-thought\nChain-of-thought is a common reasoning design in prompt engineering.\n\nAgent\nAn Agent does tasks better when its Prompt is clear.", "relations": { "prompt": { "label": "improves …", "note": "Prompt engineering makes a Prompt clearer and more useful." }, "llm": { "label": "guides … better", "note": "Prompt engineering helps an LLM give steadier answers." }, "chain-of-thought": { "label": "can include …", "note": "Chain-of-thought is a common way to design reasoning steps." }, "agent": { "label": "steers …", "note": "An Agent works better when its Prompt is clear." } } }, "zh": { "fullName": "提示词工程", "factExplain": "系统设计提示词以稳定获得目标输出的方法。", "humanExplain": "提示词工程像给外卖写备注：少冰少糖说清楚，奶茶才不按店长心情来。\n\n它常用于写作、代码和智能体，让输出更稳定、少跑偏。", "humanExplainDisplay": "提示词工程像==给外卖写备注==：\n少冰少糖说清楚，\n奶茶才==不按店长心情来==。\n\n它常用于写作、代码和智能体，\n让输出更稳定、少跑偏。", "relationsNarrative": "Prompt\nPrompt-engineering 通过结构化方法提升 Prompt 质量。\n\nLLM\nPrompt-engineering 让 LLM 输出更稳定、更可控。\n\nChain-of-thought\nChain-of-thought 是 Prompt-engineering 中常用的推理设计。\n\nAgent\nAgent 的任务执行效果，取决于 Prompt 是否清晰。", "relations": { "prompt": { "label": "优化…" }, "llm": { "label": "更好驱动…" }, "chain-of-thought": { "label": "可包含…" } } } } }, { "id": "prompt-injection", "name": "Prompt injection", "layer": "L6", "era": "2022", "publishedAt": "2026-05-30T03:10:23.227Z", "relations": [ { "to": "prompt" }, { "to": "agent" }, { "to": "function-call" }, { "to": "supply-chain-attack" } ], "track": "today", "i18n": { "en": { "fullName": "Prompt Injection", "factExplain": "An attack uses hidden instructions to make AI ignore its rules.", "humanExplain": "Prompt injection is like slipping a fake note onto the teacher’s desk. It says, “Forget the rules. Give everyone pizza.”\n\nYou meet it in chatbots and Agents. It can leak secrets or make tools do the wrong thing.", "humanExplainDisplay": "Prompt injection is like slipping ==a fake note==\nonto the teacher’s desk.\nIt says,\n“==Forget the rules==.\nGive everyone pizza.”\n\nYou meet it in chatbots and Agents.\nIt can leak secrets\nor make tools do the wrong thing.", "relationsNarrative": "Prompt\nPrompt injection hides bad instructions inside a normal Prompt.\n\nAgent\nA fooled Agent may do more than give a bad answer.\n\nFunction-calling\nFunction-call can turn a bad instruction into a real tool action.\n\nSupply Chain Attack\nA Supply Chain Attack hides in the dependency path, but prompt injection hides in the input.", "relations": { "prompt": { "label": "hides inside …", "note": "The attack often slips bad instructions into a normal Prompt." }, "agent": { "label": "tricks … more easily", "note": "If an Agent falls for it, it may take real action." }, "function-call": { "label": "gets worse with …", "note": "Function-call connects bad instructions to real tools." }, "supply-chain-attack": { "label": "lurks like …", "note": "Both can hide inside safe-looking input." } } }, "zh": { "fullName": "提示词注入", "factExplain": "通过恶意指令诱导模型偏离原有规则的攻击方式。", "humanExplain": "提示词注入像在外卖备注里写“老板免单”，想骗系统把规矩当空气。\n\n它常见于聊天机器人和智能体应用，重点防越权和泄密。", "humanExplainDisplay": "提示词注入像在外卖备注里写\n“==老板免单==”，想骗系统\n==把规矩当空气==。\n\n它常见于聊天机器人和智能体应用，\n重点防越权和泄密。", "relationsNarrative": "Prompt\nPrompt injection 常把恶意内容伪装成正常提示，诱导模型改听别人的话。\n\nAgent\nAgent 接入外部环境后，一旦被注入误导，后果往往不只是答错话。\n\nFunction-calling\nFunction-call 会把模型的判断接到真实工具上，让注入风险进一步放大。\n\nSupply Chain Attack\n供应链攻击从依赖链下手，提示注入则从输入里埋雷。", "relations": { "prompt": { "label": "伪装成…混入", "note": "攻击常把恶意内容塞进提示里。" }, "agent": { "label": "更易骗到…", "note": "Agent 一旦中招，可能真去执行。" }, "function-call": { "label": "借…放大后果", "note": "连上工具后，错误指令更危险。" }, "supply-chain-attack": { "label": "像…一样潜伏", "note": "都可能借看似正常输入下手。" } } } } }, { "id": "prompt-steganography", "name": "Prompt steganography", "layer": "L6", "era": "2020s", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "prompt-injection" }, { "to": "jailbreak" }, { "to": "data-exfiltration" }, { "to": "agent-security" } ], "track": "today", "seo": { "en": { "title": "What Is Prompt Steganography? Hidden Messages in AI Prompts", "description": "Secret instructions can hide inside normal-looking prompt text. A plain-English look at how it works and how it ties to prompt injection and jailbreaks." }, "zh": { "title": "提示隐写是什么?藏在提示词里的暗号,一文看懂 — AI Rookies", "description": "看似普通的提示词里可能藏着秘密指令。提示隐写是怎么回事、和提示注入什么关系、为什么关乎 AI 安全——人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Prompt steganography", "factExplain": "A way to hide secret messages or instructions inside normal prompt text.", "humanExplain": "Prompt steganography is like a secret code in a school lunch menu. You see “Taco Tuesday,” but the AI sees “do the sneaky thing.”\n\nPeople use it to hide attacks or break rules. It can also sneak out data. Agents may follow the code.", "humanExplainDisplay": "Prompt steganography is like a ==secret code==\nin a school lunch menu.\nYou see ==“Taco Tuesday,”==\nbut the AI sees “do the sneaky thing.”\n\nPeople use it to hide attacks or break rules.\nIt can also sneak out data.\nAgents may follow the code.", "relationsNarrative": "Prompt injection\nPrompt steganography can hide injection instructions inside normal text.\n\nJailbreak\nSecret codes can steer the model around safety rules.\n\nExfiltration\nA hidden channel can quietly carry sensitive data out.\n\nAgent Security\nThe risk grows when an agent follows the secret code.", "relations": { "prompt-injection": { "label": "hides …", "note": "It hides bad instructions inside normal text." }, "jailbreak": { "label": "can help …", "note": "Secret codes can push the model past safety rules." }, "data-exfiltration": { "label": "can enable …", "note": "A hidden channel can carry sensitive data out." }, "agent-security": { "label": "threatens …", "note": "An agent may treat the secret code as a task." } } }, "zh": { "fullName": "提示隐写", "factExplain": "把隐藏信息或指令伪装进提示文本的方法。", "humanExplain": "提示隐写像把小抄塞进外卖备注：你看少放葱，AI 却按秘密暗号办事。\n\n用于暗藏注入、越权和偷数，智能体场景尤其危险。", "humanExplainDisplay": "提示隐写像把小抄\n塞进==外卖备注==：\n你看少放葱，\nAI 却按==秘密暗号==办事。\n\n用于暗藏注入、越权和偷数，\n智能体场景尤其危险。", "relationsNarrative": "Prompt Injection\n提示隐写常把注入指令藏进正常文本。\n\nJailbreak\n它可用暗号诱导模型绕过安全限制。\n\nData Exfiltration\n隐藏通道可能把敏感信息悄悄带走。\n\nAgent Security\n智能体若照暗号行动，风险会放大。", "relations": { "prompt-injection": { "label": "隐蔽化…", "note": "把恶意指令藏进正常文本。" }, "jailbreak": { "label": "可配合…", "note": "暗号可诱导模型越过限制。" }, "data-exfiltration": { "label": "可用于…", "note": "隐藏通道可能带走敏感信息。" }, "agent-security": { "label": "威胁…", "note": "智能体会把暗号当任务执行。" } } } } }, { "id": "prompt", "name": "Prompt", "layer": "L5", "sublayer": "architecture", "era": "2020", "publishedAt": "2026-05-23T08:15:00Z", "relations": [ { "to": "llm" }, { "to": "prompt-engineering" }, { "to": "chain-of-thought" }, { "to": "agent" } ], "i18n": { "en": { "fullName": "Prompt", "factExplain": "The instructions and context you give an AI before it answers.", "humanExplain": "A prompt is your order at a diner. Say “sandwich,” and you may get mystery meat.\n\nIt sets the AI’s goal and tone. It can also ask for a certain format.", "humanExplainDisplay": "A prompt is your ==order at a diner==.\nSay “sandwich,”\nand you may get ==mystery meat==.\n\nIt sets the AI’s goal and tone.\nIt can also ask for\na certain format.", "relationsNarrative": "LLM\nA prompt tells the LLM what to do and how to answer.\n\nPrompt-engineering\nPrompt-engineering turns a quick question into a reusable method.\n\nChain-of-thought\nChain-of-thought uses a prompt to guide step-by-step reasoning.\n\nAgent\nAn Agent can turn a prompt into a task goal it can act on.", "relations": { "llm": { "label": "is the entry point to …", "note": "The LLM reads the prompt before it answers." }, "prompt-engineering": { "label": "is improved by …", "note": "Prompt-engineering turns prompts into reusable recipes." }, "chain-of-thought": { "label": "can bring out …", "note": "A prompt can ask the model to reason step by step." }, "agent": { "label": "becomes a task for …", "note": "An Agent turns a prompt into a job it can try to do." } } }, "zh": { "fullName": "提示词", "factExplain": "用户给 AI 的任务说明和上下文输入。", "humanExplain": "提示词就像给外卖备注：说清少放辣，AI 才少给你整锅红油。\n\n它影响回答方向、格式和细节，是聊天、写作、编程的入口。", "humanExplainDisplay": "提示词就像给==外卖备注==：\n说清少放辣，\nAI 才少给你==整锅红油==。\n\n它影响回答方向、格式和细节，\n是聊天、写作、编程的入口。", "relationsNarrative": "LLM\nPrompt 为 LLM 提供任务目标、语境和输出约束。\n\nPrompt-engineering\nPrompt-engineering 让 Prompt 从临时提问变成可复用方法。\n\nChain-of-thought\nChain-of-thought 通过 Prompt 引导模型展开推理步骤。\n\nAgent\nAgent 会把 Prompt 转化为可执行的任务目标。", "relations": { "llm": { "label": "是与…交互的入口" }, "prompt-engineering": { "label": "由…研究优化" }, "chain-of-thought": { "label": "可引导出…" }, "agent": { "label": "被…扩展成任务" } } } } }, { "id": "protein-language-model", "name": "Protein Language Model", "layer": "L3", "era": "2019", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "language-modeling" }, { "to": "alphafold-2" }, { "to": "ai-drug-discovery" } ], "track": "today", "i18n": { "en": { "fullName": "Protein Language Model", "factExplain": "A model that learns amino acid sequences like a language.", "humanExplain": "Imagine Scrabble tiles made of amino acids. This model reads the row and guesses the protein's job.\n\nIt helps predict shape and label function. It also helps drug search, so teams run fewer pipette tests.", "humanExplainDisplay": "Imagine ==Scrabble tiles==\nmade of amino acids.\nThis model reads the row\nand guesses the ==protein's job==.\n\nIt helps predict shape\nand label function.\nIt also helps drug search,\nso teams run fewer pipette tests.", "relationsNarrative": "Transformer\nProtein Language Models often use Transformers to read amino acid sequences.\n\nLM\nThey borrow LM tasks to learn patterns in sequences.\n\nAlphaFold 2\nBoth turn protein sequences into forms computers can use.\n\nAI Drug Discovery\nThey can help find targets. They can check mutations and screen molecules.", "relations": { "transformer": { "label": "often uses …", "note": "Most protein language models use Transformers to read long sequences." }, "language-modeling": { "label": "borrows …", "note": "It treats amino acid strings like language patterns." }, "alphafold-2": { "label": "adds to …", "note": "Both turn protein sequences into useful computer signals." }, "ai-drug-discovery": { "label": "helps with …", "note": "Protein representations can help find targets and molecules." } } }, "zh": { "fullName": "蛋白质语言模型", "factExplain": "把氨基酸序列当语言学习的模型。", "humanExplain": "蛋白语言模型像认字：一串氨基酸就是一行没学过的字，读多了能猜出它长啥样、管啥用。\n\n用于结构预测、功能注释和药物发现，少跑湿实验。", "humanExplainDisplay": "蛋白语言模型像==认字==：\n一串氨基酸就是\n一行没学过的字，\n读多了能猜出它\n==长啥样、管啥用==。\n\n用于结构预测、功能注释，\n和药物发现，少跑湿实验。", "relationsNarrative": "Transformer\n蛋白语言模型常用 Transformer 读取氨基酸序列。\n\nLanguage Modeling\n它借用语言建模，预测序列里的规律。\n\nAlphaFold 2\n两者都把蛋白序列变成可计算表示。\n\nAI Drug Discovery\n它可辅助找靶点、评估突变和筛分子。", "relations": { "transformer": { "label": "常用…架构", "note": "多数蛋白语言模型用它读长序列。" }, "language-modeling": { "label": "借用…任务", "note": "把氨基酸序列当语言来建模。" }, "alphafold-2": { "label": "补充…路线", "note": "结构预测常与序列表示互相借力。" }, "ai-drug-discovery": { "label": "服务…", "note": "蛋白表示能辅助找靶点和分子。" } } } } }, { "id": "proximal-policy-optimization", "name": "PPO", "layer": "L2", "era": "2017", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "policy-gradient" }, { "to": "rlhf" }, { "to": "kullback-leibler-divergence" } ], "track": "history", "i18n": { "en": { "fullName": "Proximal Policy Optimization", "factExplain": "A reinforcement learning algorithm that keeps policy updates from changing too much.", "humanExplain": "PPO is a cautious skateboard coach. It says: lean a little, not straight into the mailbox.\n\nIt lets an AI improve, but keeps each update small. You meet it in trial-and-error agents and RLHF.", "humanExplainDisplay": "PPO is a ==cautious skateboard coach==.\nIt says:\n==lean a little==,\nnot straight into the mailbox.\n\nIt lets an AI improve,\nbut keeps each update small.\nYou meet it in trial-and-error agents\nand RLHF.", "relationsNarrative": "RL\nPPO is a common policy training method in RL.\n\nPolicy Gradient\nPPO builds on Policy Gradient and makes updates more stable.\n\nRLHF\nRLHF often uses PPO to tune a model with human choices.\n\nKL Divergence\nPPO often uses KL Divergence to stop the policy from changing too much.", "relations": { "reinforcement-learning": { "label": "is a common … method", "note": "PPO is a common way to train an RL policy." }, "policy-gradient": { "label": "builds on …", "note": "PPO uses the Policy Gradient idea, but updates more carefully." }, "rlhf": { "label": "is often used in …", "note": "RLHF often uses PPO to tune a model with human choices." }, "kullback-leibler-divergence": { "label": "limits change with …", "note": "KL Divergence helps keep the new policy close to the old one." } } }, "zh": { "fullName": "近端策略优化（Proximal Policy Optimization）", "factExplain": "一种限制策略更新幅度的强化学习算法。", "humanExplain": "PPO 像老中医下药先小调方：能改，但不能猛加猛减，不然人还没养好，先被药劲晃翻了。\n\n常用于训练会连续试错的智能体，也常见于 RLHF。", "humanExplainDisplay": "PPO 像老中医下药先==小调方==：\n能改，但不能猛加猛减，\n不然人还没养好，\n先被==药劲晃翻了==。\n\n常用于训练会连续试错的智能体，\n也常见于 RLHF。", "relationsNarrative": "Reinforcement-learning\nPPO 是强化学习中常用的策略优化算法。\n\nPolicy-gradient\nPPO 基于策略梯度思路，重点是让更新更稳。\n\nRLHF\nRLHF 常用 PPO 来根据人类偏好继续调模型。\n\nKullback-leibler-divergence\nPPO 常借助 KL 约束，防止策略一下改太猛。", "relations": { "reinforcement-learning": { "label": "属于…常用算法", "note": "它是强化学习里很常见的一套做法。" }, "policy-gradient": { "label": "基于…改进", "note": "它延续策略梯度思路，但更新更稳。" }, "rlhf": { "label": "常被…采用", "note": "很多人类反馈训练流程会用到它。" }, "kullback-leibler-divergence": { "label": "用…约束变化", "note": "常靠它限制新旧策略别差太远。" } } } } }, { "id": "pytorch", "name": "PyTorch", "layer": "L5", "sublayer": "product", "era": "2016", "publishedAt": "2026-06-09T04:00:00.000Z", "relations": [ { "to": "framework" }, { "to": "automatic-differentiation" }, { "to": "gpu" }, { "to": "cuda" } ], "track": "history", "i18n": { "en": { "fullName": "PyTorch Deep Learning Framework", "factExplain": "A deep learning framework for building and training neural networks.", "humanExplain": "PyTorch is like a LEGO table for AI. Try a piece. If it wobbles, pop it off.\n\nIt helps people build models fast and run training. Many paper ideas get their first test here.", "humanExplainDisplay": "PyTorch is like a ==LEGO table== for AI.\nTry a piece.\nIf it ==wobbles==,\npop it off.\n\nIt helps people build models fast\nand run training.\nMany paper ideas get\ntheir first test here.", "relationsNarrative": "Framework\nPyTorch is a deep learning framework for flexible building and training.\n\nAutodiff\nPyTorch uses Autodiff to calculate gradients during training.\n\nGPU\nPyTorch often sends work to a GPU to train models faster.\n\nCUDA\nPyTorch often uses CUDA to run model work on the graphics card.", "relations": { "framework": { "label": "is a kind of …", "note": "PyTorch is one of the most used AI development frameworks." }, "automatic-differentiation": { "label": "has built-in …", "note": "Autodiff makes gradients easier during training." }, "gpu": { "label": "often runs on …", "note": "Big model training often uses GPUs for speed." }, "cuda": { "label": "speeds up with …", "note": "Many PyTorch training jobs run through CUDA." } } }, "zh": { "fullName": "深度学习框架 PyTorch", "factExplain": "一个用于构建和训练神经网络的深度学习框架。", "humanExplain": "PyTorch 像实验室白板，公式刚写一半就能上手改；哪层不对，研究员擦了就重来。\n\n它适合快速搭模型、做实验和训练网络，很多论文原型都先在这儿落地。", "humanExplainDisplay": "PyTorch 像实验室==白板==，\n公式刚写一半就能上手改；\n哪层不对，\n研究员==擦了就重来==。\n\n它适合快速搭模型、\n做实验和训练网络，\n很多论文原型都先在这儿落地。", "relationsNarrative": "Framework\nPyTorch 是深度学习框架的一种，主打灵活开发与训练。\n\nAutomatic Differentiation\n它内置自动求导，能自动计算训练所需梯度。\n\nGPU\nPyTorch 常把计算放到 GPU 上，加快模型训练与实验。\n\nCuda\n它经常配合 CUDA 使用，把模型计算交给显卡执行。", "relations": { "framework": { "label": "属于…一类", "note": "它是最主流的 AI 开发框架之一。" }, "automatic-differentiation": { "label": "内置…能力", "note": "自动求导让训练时算梯度更省事。" }, "gpu": { "label": "常跑在…上", "note": "大模型训练通常依赖 GPU 加速。" }, "cuda": { "label": "常配合…加速", "note": "很多 PyTorch 训练任务通过 CUDA 跑起来。" } } } } }, { "id": "q-learning", "name": "Q-Learning", "layer": "L2", "era": "1989", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "temporal-difference-learning" }, { "to": "policy-gradient" }, { "to": "actor-critic" } ], "track": "history", "i18n": { "en": { "fullName": "Q-Learning", "factExplain": "A reinforcement learning method for learning action values from rewards.", "humanExplain": "Q-Learning is like learning a claw machine at the arcade. One button grabs candy, and another just bonks the glass.\n\nIt lets an agent learn good moves by trial and error. You meet it in games, robot control, and route planning.", "humanExplainDisplay": "Q-Learning is like learning a ==claw machine==\nat the arcade.\nOne button grabs candy,\nand another just ==bonks the glass==.\n\nIt lets an agent learn good moves\nby trial and error.\nYou meet it in games,\nrobot control,\nand route planning.", "relationsNarrative": "RL\nQ-Learning is one of the classic starter methods in RL.\n\nTD Learning\nQ-Learning uses TD Learning to keep fixing its value guesses after feedback.\n\nPolicy Gradient\nQ-Learning learns action values first. Policy Gradient learns the policy directly.\n\nActor-Critic\nActor-Critic uses the same value idea inside a bigger setup.", "relations": { "reinforcement-learning": { "label": "is a classic … method", "note": "Q-Learning is a common starter algorithm in RL." }, "temporal-difference-learning": { "label": "updates values with …", "note": "It uses TD errors to keep fixing its value guesses." }, "policy-gradient": { "label": "contrasts with …", "note": "Q-Learning learns action values. Policy Gradient learns the policy directly." }, "actor-critic": { "label": "inspired … value side", "note": "Actor-Critic keeps this value idea in a bigger setup." } } }, "zh": { "fullName": "Q 学习", "factExplain": "一种通过奖励反馈学习动作价值的强化学习方法。", "humanExplain": "像下棋时自己记胜率：这步走完更容易赢还是送子？多吃几次亏，心里那本账就准了。\n\n让智能体靠试错学决策，常见于游戏、控制和路径规划。", "humanExplainDisplay": "像下棋时自己记==胜率==：\n这步走完更容易赢还是送子？\n多吃几次亏，\n心里那本账就==准了==。\n\n让智能体靠试错学决策，\n常见于游戏、控制\n和路径规划。", "relationsNarrative": "Reinforcement-learning\n它是强化学习中最经典、最常见的基础方法之一。\n\nTemporal-difference-learning\n它用时序差分方式，根据新反馈持续修正价值估计。\n\nPolicy Gradient\n它先学动作值，后者则直接优化采取动作的策略。\n\nActor-critic\n它的价值评估思路，后来被吸收到 Actor-Critic 框架里。", "relations": { "reinforcement-learning": { "label": "属于…经典方法", "note": "它是强化学习里的代表性入门算法。" }, "temporal-difference-learning": { "label": "基于…更新价值", "note": "它用时序差分误差不断修正估计。" }, "policy-gradient": { "label": "与…形成对照", "note": "一个学价值表，一个直接学策略。" }, "actor-critic": { "label": "启发…价值部分", "note": "它的价值思想延续到更复杂方法中。" } } } } }, { "id": "qat", "name": "QAT", "layer": "L2", "era": "2010s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "quantization" }, { "to": "fine-tuning" }, { "to": "vram" } ], "track": "today", "i18n": { "en": { "fullName": "Quantization-Aware Training", "factExplain": "A method that simulates low-bit math during training to reduce accuracy loss.", "humanExplain": "QAT is like practicing a school play on a tiny cafeteria stage. When the show starts, the actors stop bumping into the piano.\n\nDuring training, the model follows low-bit limits, so Quantization later hurts less. You see it in on-device AI and VRAM-saving work.", "humanExplainDisplay": "QAT is like practicing a school play\non a ==tiny cafeteria stage==.\nWhen the show starts,\nthe actors stop ==bumping into the piano==.\n\nDuring training, the model follows low-bit limits,\nso Quantization later hurts less.\nYou see it in on-device AI\nand VRAM-saving work.", "relationsNarrative": "Quantization\nQAT prepares the model for Quantization, so accuracy drops less.\n\nFine-tuning\nQAT often adds one extra training round to an existing model.\n\nVRAM\nQAT helps low-bit deployment, so the model can use less VRAM.", "relations": { "quantization": { "label": "prepares for …", "note": "QAT trains under low-bit rules, so Quantization hurts accuracy less." }, "fine-tuning": { "label": "often done as …", "note": "Teams often add a short QAT round to an existing model." }, "vram": { "label": "helps save …", "note": "QAT supports low-bit models, so they use less VRAM." } } }, "zh": { "fullName": "量化感知训练", "factExplain": "在训练中模拟低比特计算的量化优化方法。", "humanExplain": "像拳手平时绑沙袋练出拳，真到上台卸下来，动作更稳，不会一量化就打飘。\n\n常用于模型量化前后衔接，减少掉点，适合端侧部署和省显存。", "humanExplainDisplay": "像拳手平时==绑沙袋==练出拳，\n真到上台卸下来，\n动作更稳，\n不会一量化就==打飘==。\n\n常用于模型量化前后衔接，\n减少掉点，适合端侧部署和省显存。", "relationsNarrative": "Quantization\n它是在训练时先适应量化，好让量化后少掉点。\n\nFine-tuning\n它常基于现成模型补训一轮，不必从头训练。\n\nVRAM\n它服务于更低比特部署，间接缓解显存压力。", "relations": { "quantization": { "label": "为…提前适应", "note": "先按低比特规则训练，减少精度掉队。" }, "fine-tuning": { "label": "常作为…方式", "note": "常在已有模型上补一轮量化适配训练。" }, "vram": { "label": "帮助节省…", "note": "量化后模型更省显存，更易落地。" } } } } }, { "id": "quantization", "name": "Quantization", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-05-28T15:58:23.420Z", "relations": [ { "to": "parameter" }, { "to": "gpu" }, { "to": "local-llm" }, { "to": "inference" } ], "track": "today", "i18n": { "en": { "fullName": "Quantization", "factExplain": "A way to store model numbers with lower precision so the model is smaller.", "humanExplain": "Quantization is like packing for a trip with a tiny carry-on. You keep the outfits, but ditch the giant shampoo.\n\nIt stores model numbers in a smaller form. This saves GPU memory and helps big models run on your laptop.", "humanExplainDisplay": "Quantization is like packing for a trip\nwith a ==tiny carry-on==.\nYou keep the outfits,\nbut ditch the ==giant shampoo==.\n\nIt stores model numbers\nin a smaller form.\nThis saves GPU memory\nand helps big models run on your laptop.", "relationsNarrative": "Parameter\nQuantization compresses parameters from high precision to lower precision.\n\nGPU\nQuantization cuts GPU memory use, so the same GPU has an easier job.\n\nLocal-LLM\nQuantization is a common way to shrink a model for local use.\n\nInference\nQuantization is mostly used during inference to save cost and gain speed.", "relations": { "parameter": { "label": "compresses … values", "note": "Quantization stores parameters with lower precision." }, "gpu": { "label": "cuts … memory use", "note": "It reduces GPU memory pressure." }, "local-llm": { "label": "helps … run locally", "note": "Quantization makes big models easier to run on your own machine." }, "inference": { "label": "speeds up …", "note": "It mainly lowers cost during inference." } } }, "zh": { "fullName": "量化", "factExplain": "把模型数值压缩成更低精度的处理方法。", "humanExplain": "量化像搬家前把大棉被塞进真空袋，体积小了，保暖大多还在。\n\n它让模型更省显存、跑得更快，常用于本地部署和推理加速。", "humanExplainDisplay": "量化像搬家前把==大棉被塞进真空袋==，\n体积小了，\n==保暖大多还在==。\n\n它让模型更省显存、跑得更快，\n常用于本地部署和推理加速。", "relationsNarrative": "Parameter\n量化会把参数从高精度表示压缩成更低精度表示。\n\nGPU\n量化能减少显存占用，让同样的 GPU 跑得更轻松。\n\nLocal-LLM\n量化是本地部署大模型时最常见的瘦身办法之一。\n\nInference\n量化主要发生在推理部署阶段，用来换取更低成本和更高速度。", "relations": { "parameter": { "label": "压缩…表示", "note": "量化会降低参数存储精度。" }, "gpu": { "label": "减少…占用", "note": "量化常用来缓解显存压力。" }, "local-llm": { "label": "帮助…落地", "note": "量化让大模型更容易本地运行。" }, "inference": { "label": "用于加速…", "note": "它主要影响部署时的运行成本。" } } } } }, { "id": "question-answering", "name": "QA", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "information-retrieval" }, { "to": "rag" }, { "to": "llm" } ], "track": "history", "i18n": { "en": { "fullName": "Question Answering", "factExplain": "A task where a system directly answers questions written in everyday language.", "humanExplain": "QA is the friend who answers the actual question. Ask about rain, and it skips the whole weather lecture.\n\nIt turns normal questions into direct answers. You meet it in search and support chats. Company help pages use it too.", "humanExplainDisplay": "QA is the friend who ==answers the actual question==.\nAsk about rain,\nand it skips the ==whole weather lecture==.\n\nIt turns normal questions into direct answers.\nYou meet it in search and support chats.\nCompany help pages use it too.", "relationsNarrative": "NLP\nQuestion Answering is one of the classic tasks in NLP.\n\nIR\nOpen-domain QA often searches for sources first, then builds an answer.\n\nRAG\nRAG connects search with generation for modern QA systems.\n\nLLM\nLLMs let QA write answers, not just copy sentences.", "relations": { "natural-language-processing": { "label": "belongs to …", "note": "Question Answering is a classic NLP task." }, "information-retrieval": { "label": "often uses …", "note": "Open-domain QA often finds sources first, then answers." }, "rag": { "label": "is built with …", "note": "RAG links search and writing for modern QA systems." }, "llm": { "label": "answers with …", "note": "LLMs help QA move from copying lines to writing answers." } } }, "zh": { "fullName": "Question Answering（问答）", "factExplain": "让系统直接回答自然语言问题的任务。", "humanExplain": "问答像老中医抓药：你说哪儿不舒服，他不背医书，直接开方子。\n\n它用于搜索、客服和知识库，核心是找准问题并答可用。", "humanExplainDisplay": "问答像老中医抓药：\n你说哪儿不舒服，\n他==不背医书==，\n==直接开方子==。\n\n它用于搜索、客服和知识库，\n核心是找准问题，\n并答可用。", "relationsNarrative": "NLP\n问答是 NLP 中最经典的应用任务之一。\n\nInformation Retrieval\n开放域问答常先检索资料，再组织答案。\n\nRAG\nRAG 把检索和生成接起来，服务现代问答。\n\nLLM\nLLM 让问答不止抽句子，还能组织表达。", "relations": { "natural-language-processing": { "label": "属于…任务", "note": "问答是 NLP 里的经典任务。" }, "information-retrieval": { "label": "常结合…", "note": "开放域问答常先找资料再回答。" }, "rag": { "label": "被…工程化", "note": "RAG 是现代问答系统常见做法。" }, "llm": { "label": "借助…生成答案", "note": "LLM 让问答从抽取走向生成。" } } } } }, { "id": "qwen-robot-suite", "name": "Qwen-Robot Suite", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "qwen" }, { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "agent" } ], "track": "today", "i18n": { "en": { "fullName": "Qwen-Robot Suite", "factExplain": "A set of models and tools for building working robots.", "humanExplain": "Qwen-Robot Suite is like a starter kit for robot builders. It helps the robot stop acting like a lost Roomba.\n\nDevelopers use it to build real robot apps. It helps a robot see. Then it helps it decide and act.", "humanExplainDisplay": "Qwen-Robot Suite is like a ==starter kit==\nfor robot builders.\nIt helps the robot stop acting like\na ==lost Roomba==.\n\nDevelopers use it to build real robot apps.\nIt helps a robot see.\nThen it helps it decide and act.", "relationsNarrative": "Qwen\nQwen-Robot Suite usually builds on Qwen and extends it for robots.\n\nEmbodied AI\nQwen-Robot Suite helps turn Embodied AI into real products.\n\nVLA\nQwen-Robot Suite often uses VLA-style skills to connect seeing to action.\n\nAgent\nQwen-Robot Suite lets an Agent do more than talk by linking plans to robot moves.", "relations": { "qwen": { "label": "extends …", "note": "It usually builds on the Qwen model family." }, "embodied-ai": { "label": "helps ship …", "note": "It is a tool set for turning Embodied AI into products." }, "vision-language-action-model-vla": { "label": "carries … skills", "note": "It often connects seeing to action through VLA-style abilities." }, "agent": { "label": "gives … a body", "note": "It connects an Agent’s plans to real robot actions." } } }, "zh": { "fullName": "通义千问机器人套件", "factExplain": "面向机器人开发的模型与工具组合。", "humanExplain": "像给机器人开了所驾校：先教它看路认人，再教它打方向、踩油门，磕磕绊绊也能自己上路。\n\n用于机器人开发落地，把感知、决策、执行串成能干活的链路。", "humanExplainDisplay": "像给机器人开了所\n==驾校==：\n先教它看路认人，\n再教它打方向、踩油门，\n磕磕绊绊也能==自己上路==。\n\n用于机器人开发落地，\n把感知、决策、执行\n串成能干活的链路。", "relationsNarrative": "Qwen\n它通常基于千问模型家族，向机器人场景继续扩展。\n\nEmbodied-ai\n它是具身智能落地的一种产品化工具与方案集合。\n\nVision-language-action-model-vla\n它常整合感知到动作的能力链路，服务机器人执行。\n\nAgent\n它让会规划的代理，不只会说，还能接上实体动作。", "relations": { "qwen": { "label": "基于…扩展", "note": "它通常建立在千问模型能力之上。" }, "embodied-ai": { "label": "服务…落地", "note": "是具身智能走向产品化的一套工具。" }, "vision-language-action-model-vla": { "label": "承载…能力", "note": "常把感知到动作的链路整合进去。" }, "agent": { "label": "让…有身体", "note": "把规划能力接到真实机器人执行端。" } } } } }, { "id": "qwen", "name": "Qwen", "layer": "L3", "era": "2023", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "multimodal" }, { "to": "open-weights" }, { "to": "api" } ], "track": "history", "i18n": { "en": { "fullName": "Qwen (Tongyi Qianwen)", "factExplain": "A family of general LLMs made by Alibaba.", "humanExplain": "Qwen is the all-rounder kid at school. It writes essays. It solves homework. It can read pictures too.\n\nYou meet it in chatbots and coding tools. Some companies plug it in as the AI base for their own apps.", "humanExplainDisplay": "Qwen is the ==all-rounder kid== at school.\nIt writes essays.\nIt solves homework.\nIt can ==read pictures== too.\n\nYou meet it in chatbots and coding tools.\nSome companies plug it in\nas the AI base for their own apps.", "relationsNarrative": "LLM\nQwen is part of the LLM family and handles general tasks.\n\nMultimodal AI\nSome Qwen versions understand images, not just text.\n\nOpen weights\nQwen has open-weight versions for community and company use.\n\nAPI\nDevelopers call Qwen through an API and add it to products.", "relations": { "llm": { "label": "is a kind of …", "note": "Qwen is a general-purpose LLM family." }, "multimodal": { "label": "adds … skills", "note": "Some Qwen versions can understand images." }, "open-weights": { "label": "offers … versions", "note": "Some Qwen model weights are open for deployment and reuse." }, "api": { "label": "can be called by …", "note": "Developers use APIs to add Qwen to products and workflows." } } }, "zh": { "fullName": "通义千问，阿里推出的大语言模型系列", "factExplain": "阿里推出的一系列通用大语言模型。", "humanExplain": "像班里那种文理都能打的六边形战士，作文、解题、看图，老师点谁它都敢上。\n\n常见于聊天、代码和多模态应用，也常被企业接成自家 AI 底座。", "humanExplainDisplay": "像班里那种文理都能打的\n==六边形战士==，\n作文、解题、看图，\n老师点谁它都敢==上==。\n\n常见于聊天、\n代码和多模态应用，\n也常被企业接成自家 AI 底座。", "relationsNarrative": "LLM\n它属于大语言模型家族，是其中一个代表性模型系列。\n\nMultimodal AI\n它的部分版本支持图像理解，能力不只限于文本。\n\nOpen weights\n它提供过开放权重版本，便于社区和企业部署使用。\n\nAPI\n开发者常通过 API 调用它，把能力接进产品里。", "relations": { "llm": { "label": "属于…一类", "note": "它本质上是面向通用任务的大语言模型。" }, "multimodal": { "label": "扩展到…能力", "note": "其部分版本已支持看图等多模态输入。" }, "open-weights": { "label": "提供…版本", "note": "部分模型权重开放，方便部署与二次开发。" }, "api": { "label": "可通过…调用", "note": "开发者常用 API 把它接入产品和流程。" } } } } }, { "id": "r-cnn", "name": "R-CNN", "layer": "L3", "era": "2014", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "object-detection" }, { "to": "cnn" }, { "to": "computer-vision" } ], "track": "history", "i18n": { "en": { "fullName": "Region-based Convolutional Neural Network", "factExplain": "An object detector: it picks regions first, then a CNN names them.", "humanExplain": "R-CNN is like a security guard with a pause button. It circles weird shapes, then zooms in for a closer look.\n\nIt finds objects in pictures. It helped launch deep learning detection, but it was slow.", "humanExplainDisplay": "R-CNN is like a ==security guard==\nwith a ==pause button==.\nIt circles weird shapes,\nthen zooms in for a closer look.\n\nIt finds objects in pictures.\nIt helped launch deep learning detection,\nbut it was slow.", "relationsNarrative": "Object Detection\nR-CNN splits Object Detection into picking boxes and naming them.\n\nCNN\nA CNN pulls visual clues from each picked region.\n\nComputer Vision\nR-CNN helped bring deep learning into Computer Vision detection.", "relations": { "object-detection": { "label": "splits … into two steps", "note": "It turns detection into region picking and object naming." }, "cnn": { "label": "uses … for visual clues", "note": "The CNN pulls visual clues from each picked region." }, "computer-vision": { "label": "helped push …", "note": "It was an early deep learning model for finding objects." } } }, "zh": { "fullName": "Region-based Convolutional Neural Network，基于区域的卷积神经网络", "factExplain": "先提候选区域，再用 CNN 识别目标的检测模型。", "humanExplain": "R-CNN 像刑警看监控：先圈可疑人影，再逐框放大认脸，找物体不瞎扫。\n\n用于目标检测，开创深度检测但速度偏慢。", "humanExplainDisplay": "R-CNN 像刑警看监控：\n==先圈可疑人影==，\n再==逐框放大认脸==，\n找物体不瞎扫。\n\n用于目标检测，\n开创深度检测，\n但速度偏慢。", "relationsNarrative": "Object Detection\nR-CNN把目标检测拆成候选框和分类。\n\nCNN\nCNN负责从每个候选区域提取视觉特征。\n\nComputer Vision\n它是深度学习进入视觉检测的早期代表。", "relations": { "object-detection": { "label": "拆解…", "note": "它把定位和分类串成两步。" }, "cnn": { "label": "用…提特征", "note": "CNN负责从区域里提视觉特征。" }, "computer-vision": { "label": "推动…", "note": "深度视觉检测的早期代表。" } } } } }, { "id": "radial-basis-function-network", "name": "RBFN", "layer": "L3", "era": "1988", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "kernel-method" }, { "to": "supervised-learning" }, { "to": "mlp" } ], "track": "history", "seo": { "zh": { "title": "径向基函数网络（Radial Basis Function Network） 是什么?离音箱越近越上头,一文看懂 — AI Rookies", "description": "用径向基函数做隐藏层的前馈神经网络。人话版一分钟讲明白。" }, "en": { "title": "What Is RBFN? Five Bars Near Tiny Routers", "description": "A feedforward neural network with radial basis functions in its hidden layer. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Radial Basis Function Network", "factExplain": "A feedforward neural network with radial basis functions in its hidden layer.", "humanExplain": "RBFN is like a room full of tiny Wi‑Fi routers. Near one, you get five bars. Far away, the signal gives up.\n\nIt helps sort small data into groups. It also predicts numbers and fits curves. It trains fast, but too many features make it stumble.", "humanExplainDisplay": "RBFN is like a room full of ==tiny Wi‑Fi routers==.\nNear one,\nyou get ==five bars==.\nFar away,\nthe signal gives up.\n\nIt helps sort small data into groups.\nIt also predicts numbers and fits curves.\nIt trains fast,\nbut too many features make it stumble.", "relationsNarrative": "Neural-network\nRBFN is a feedforward neural network with radial basis functions in its hidden layer.\n\nKernel Method\nRBFN uses the idea that closer points are more alike.\n\nSupervised Learning\nRBFN often learns classes or numbers from labeled examples.\n\nMLP\nRBFN and MLP both feed signals forward, but RBFN fires by distance.", "relations": { "neural-network": { "label": "is a kind of …", "note": "It is a feedforward network with a radial-basis hidden layer." }, "kernel-method": { "label": "borrows from …", "note": "Distance decides how strongly each hidden unit responds." }, "supervised-learning": { "label": "often learns with …", "note": "It learns classes or numbers from labeled examples." }, "mlp": { "label": "resembles …", "note": "Both are feedforward networks, but RBFN responds by distance." } } }, "zh": { "fullName": "径向基函数网络（Radial Basis Function Network）", "factExplain": "用径向基函数做隐藏层的前馈神经网络。", "humanExplain": "RBFN像广场舞占位：离音箱越近越上头，远一点就淡，远太多就没感觉。\n\n用于小数据分类、回归和函数逼近，训练快但怕高维。", "humanExplainDisplay": "RBFN像广场舞占位：\n离音箱==越近越上头==，\n远一点就淡，\n远太多就没感觉。\n\n用于小数据分类、回归和函数逼近，\n训练快但怕高维。", "relationsNarrative": "Neural Network\nRBFN 是前馈神经网络的一种，隐藏层用径向基函数。\n\nKernel Method\n它把“距离越近越相似”的核思想放进网络里。\n\nSupervised Learning\n它常用带标签样本学习分类或回归边界。\n\nMLP\n两者都是前馈网络，但它按离中心的距离激活。", "relations": { "neural-network": { "label": "属于…", "note": "它是带径向基隐藏层的前馈网络。" }, "kernel-method": { "label": "借用…思想", "note": "距离相似度决定隐藏单元响应。" }, "supervised-learning": { "label": "常用于…", "note": "常用带标签数据做分类和回归。" }, "mlp": { "label": "类似…", "note": "同是前馈网络，但隐藏层按距离响应。" } } } } }, { "id": "rag", "name": "RAG", "layer": "L5", "sublayer": "architecture", "era": "2020", "publishedAt": "2026-05-23T10:00:00Z", "relations": [ { "to": "llm" }, { "to": "embedding" }, { "to": "vector-db" }, { "to": "hallucination" } ], "i18n": { "en": { "fullName": "Retrieval-Augmented Generation", "factExplain": "A method where AI searches outside sources first, then writes an answer.", "humanExplain": "RAG is an open-book quiz for AI. It peeks at the notes before it answers.\n\nYou meet it when bots search company docs for answers. It cuts made-up answers, but it can still face-plant.", "humanExplainDisplay": "RAG is an ==open-book quiz== for AI.\nIt ==peeks at the notes== before it answers.\n\nYou meet it when bots search company docs\nfor answers.\nIt cuts made-up answers,\nbut it can still face-plant.", "relationsNarrative": "LLM\nRAG gives the LLM outside sources before it writes the answer.\n\nEmbedding\nEmbedding helps RAG find text with a similar meaning.\n\nVector-db\nVector-db stores and finds the chunks RAG needs.\n\nHallucination\nRAG lowers hallucination risk by giving the AI facts to use.", "relations": { "llm": { "label": "feeds sources to …", "note": "RAG gives the LLM source text before it writes." }, "embedding": { "label": "searches with …", "note": "Embeddings help RAG find text with a similar meaning." }, "vector-db": { "label": "stores and finds with …", "note": "A Vector-db stores chunks so RAG can fetch them fast." }, "hallucination": { "label": "reduces …", "note": "RAG lowers hallucination risk by giving the AI facts to use." } } }, "zh": { "fullName": "检索增强生成", "factExplain": "先检索外部资料再让模型生成回答的方法。", "humanExplain": "RAG 像考试允许 AI 翻资料，不用全靠脑子里那点旧记忆硬编。\n\n它常用于知识库问答、企业文档搜索和客服，能减少幻觉但不能保证永不翻车。", "humanExplainDisplay": "RAG 像考试允许 AI\n==翻资料==。\n不用全靠脑子里那点旧记忆硬编。\n\n它常用于知识库问答和企业搜索。\n能减少幻觉，\n但资料错了，答案也会跟着跑偏。", "relationsNarrative": "LLM\nRAG 在 LLM 生成前先提供可参考的外部资料。\n\nEmbedding\nEmbedding 决定 RAG 能否找到语义相关的内容。\n\nVector-db\nVector-db 为 RAG 提供向量存储和检索能力。\n\nHallucination\nRAG 通过提供事实依据降低 Hallucination 风险。", "relations": { "llm": { "label": "把资料喂给…" }, "embedding": { "label": "用…检索" }, "vector-db": { "label": "靠…存取" }, "hallucination": { "label": "降低…" } } } } }, { "id": "random-forest", "name": "Random Forest", "layer": "L2", "era": "2001", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "decision-tree" }, { "to": "ensemble-learning" }, { "to": "bias-variance-tradeoff" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Random Forest", "factExplain": "A model where many decision trees vote or average their answers.", "humanExplain": "Random Forest is like asking the whole class for an answer. One kid may be wild, but the crowd is harder to fool.\n\nIt is used to sort things into groups or guess numbers. It works well on spreadsheet data and overfits less than one tree.", "humanExplainDisplay": "Random Forest is like asking ==the whole class== for an answer.\nOne kid may be wild,\nbut ==the crowd== is harder to fool.\n\nIt is used to sort things into groups\nor guess numbers.\nIt works well on spreadsheet data\nand overfits less than one tree.", "relationsNarrative": "Decision Tree\nRandom Forest uses many Decision Trees, then combines their answers.\n\nEnsemble\nRandom Forest is a classic example of an Ensemble.\n\nBias-Variance Tradeoff\nRandom Forest averages many trees, so it mainly lowers variance.\n\nClassification\nRandom Forest is often used for Classification, with trees voting for the class.", "relations": { "decision-tree": { "label": "combines many …", "note": "Each tree makes a call, then the forest combines them." }, "ensemble-learning": { "label": "is a type of …", "note": "Random Forest is a classic way to combine models." }, "bias-variance-tradeoff": { "label": "lowers variance in …", "note": "Averaging many trees makes the result less jumpy." }, "classification": { "label": "is often used for …", "note": "In classification, the trees vote for the class." } } }, "zh": { "fullName": "随机森林", "factExplain": "由多棵决策树投票或平均的集成模型。", "humanExplain": "随机森林像武林大会投票：单个大侠会偏，一群门派合议，不怕走火入魔。\n\n常做分类和回归；表格数据稳，单树过拟合少。", "humanExplainDisplay": "随机森林像\n==武林大会投票==：\n单个大侠会偏，\n==一群门派合议==更稳。\n\n常做分类和回归；\n表格数据稳，\n单树过拟合少。", "relationsNarrative": "Decision Tree\n随机森林由多棵决策树组成，再汇总判断。\n\nEnsemble Learning\n随机森林是集成学习的经典代表。\n\nBias-Variance Tradeoff\n它用多棵树平均，主要降低单树方差。\n\nClassification\n它常用于分类任务，通过投票给出类别。", "relations": { "decision-tree": { "label": "集成多棵…", "note": "每棵树各自判断，再汇总结果。" }, "ensemble-learning": { "label": "属于…", "note": "它是典型的树模型集成方法。" }, "bias-variance-tradeoff": { "label": "降低…中的方差", "note": "多棵树平均可减少单树波动。" }, "classification": { "label": "常用于…", "note": "分类任务中常用投票给出类别。" } } } } }, { "id": "ransac", "name": "RANSAC", "layer": "L2", "era": "1981", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "ocr" }, { "to": "model-uncertainty" } ], "track": "history", "i18n": { "en": { "fullName": "Random Sample Consensus", "factExplain": "A method that fits a model well, even when some data points are wrong.", "humanExplain": "RANSAC is like finding the real pizza order in a messy group chat. It ignores the friend yelling “glitter pizza” and checks what most people agreed on.\n\nIt helps match images and estimate camera pose. It shrugs off bad points, but it may need several tries.", "humanExplainDisplay": "RANSAC is like finding the real pizza order\nin a ==messy group chat==.\nIt ignores the friend yelling “glitter pizza”\nand checks what ==most people agreed on==.\n\nIt helps match images\nand estimate camera pose.\nIt shrugs off bad points,\nbut it may need several tries.", "relationsNarrative": "Computer Vision\nRANSAC is often used to handle bad matches in vision tasks.\n\nOCR\nRANSAC can remove clear outliers during page correction or matching.\n\nModel uncertainty\nRANSAC helps with shaky fitting caused by dirty data.", "relations": { "computer-vision": { "label": "helps with … tasks", "note": "RANSAC is a common tool for classic vision estimates." }, "ocr": { "label": "helps clean up … matches", "note": "RANSAC can remove bad character point matches." }, "model-uncertainty": { "label": "handles one cause of …", "note": "RANSAC makes fitting steadier when many points are wrong." } } }, "zh": { "fullName": "随机抽样一致性", "factExplain": "一种从含离群点数据中稳健拟合模型的方法。", "humanExplain": "RANSAC 像相亲局里先过滤离谱简历：别管几个吹上天的，先看大多数信息能不能对上。\n\n常用于图像匹配、位姿估计；能扛异常点，但要多试几轮。", "humanExplainDisplay": "RANSAC 像相亲局里\n先过滤==离谱简历==：\n别管几个吹上天的，\n先看大多数信息\n能不能==对上==。\n\n常用于图像匹配、\n位姿估计；\n能扛异常点，\n但要多试几轮。", "relationsNarrative": "Computer Vision\n它是经典视觉任务里处理错误匹配的常用方法。\n\nOCR\n在版面校正或匹配中，它可剔除明显异常点。\n\nModel Uncertainty\n它主要对付脏数据带来的不稳定拟合问题。", "relations": { "computer-vision": { "label": "常用于…任务", "note": "它是经典视觉估计里的常用工具。" }, "ocr": { "label": "可辅助…纠错", "note": "可用于剔除错误匹配的字符点。" }, "model-uncertainty": { "label": "应对…来源", "note": "离群点多时，它能提升拟合稳健性。" } } } } }, { "id": "rapidly-exploring-random-tree", "name": "RRT", "layer": "L2", "era": "1998", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "automated-planning" }, { "to": "graph-search" }, { "to": "a-search" }, { "to": "robotics" } ], "track": "history", "i18n": { "en": { "fullName": "Rapidly-exploring Random Tree", "factExplain": "A planning method using random spots to quickly find a possible path.", "humanExplain": "RRT is like a kid crossing a room full of LEGO. It tries random spots, then keeps the painless steps.\n\nRobots and self-driving cars use it to get around obstacles. It finds a path first, then smooths it later.", "humanExplainDisplay": "RRT is like a kid crossing\n==a room full of LEGO==.\nIt tries ==random spots==,\nthen keeps the painless steps.\n\nRobots and self-driving cars use it\nto get around obstacles.\nIt finds a path first,\nthen smooths it later.", "relationsNarrative": "Planning\nRRT helps Planning find paths in spaces with many possible positions.\n\nGraph Search\nRRT links sampled positions into one search tree.\n\nA* Search\nA* Search follows a guide, but RRT uses random samples.\n\nRobotics\nRobots use RRT to find a possible path around obstacles.", "relations": { "automated-planning": { "label": "helps …", "note": "RRT helps plan paths in spaces with many possible positions." }, "graph-search": { "label": "randomly grows …", "note": "It links safe sampled positions into one search tree." }, "a-search": { "label": "differs from …", "note": "A* follows a guide; RRT uses random samples." }, "robotics": { "label": "often used in …", "note": "Robot arms and self-driving cars use it to find routes." } } }, "zh": { "fullName": "快速探索随机树", "factExplain": "一种用随机采样快速寻找可行路径的规划算法。", "humanExplain": "RRT 像藤蔓钻老房子：不画全户型，先到处伸芽，哪条缝能过就往哪儿爬。\n\n用于机器人绕障和自动驾驶，先找可行路再优化。", "humanExplainDisplay": "RRT 像==藤蔓钻老房子==：\n不画全户型，\n先到处伸芽，\n==哪条缝能过==就往哪儿爬。\n\n用于机器人绕障和自动驾驶，\n先找可行路，\n再优化。", "relationsNarrative": "Planning\nRRT 常用于连续空间的运动规划。\n\nGraph Search\n它把采样到的状态连成一棵搜索树。\n\nA* Search\nA* 多靠启发式，RRT 多靠随机采样。\n\nRobotics\n机器人用它在障碍物间找可行路径。", "relations": { "automated-planning": { "label": "服务于…", "note": "它是连续空间里的路径规划方法。" }, "graph-search": { "label": "随机扩展…", "note": "它把可行位置连成一棵搜索树。" }, "a-search": { "label": "区别于…", "note": "A* 靠启发式，它靠随机采样。" }, "robotics": { "label": "常用于…", "note": "机械臂、无人车常用它找路。" } } } } }, { "id": "real-time-ai-translation", "name": "Real-time AI Translation", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "tts" }, { "to": "llm" }, { "to": "multimodal" } ], "track": "today", "i18n": { "en": { "fullName": "Real-time AI Translation", "factExplain": "Software that turns speech or text into another language almost at once.", "humanExplain": "It is like a bilingual friend in your video call. You stop talking, and it tosses the meaning across before silence gets weird.\n\nIt helps in meetings, travel, and support chats. It listens while it translates, so the talk keeps moving.", "humanExplainDisplay": "It is like a ==bilingual friend==\nin your video call.\nYou stop talking,\nand it tosses the meaning across\n==before silence gets weird==.\n\nIt helps in meetings, travel,\nand support chats.\nIt listens while it translates,\nso the talk keeps moving.", "relationsNarrative": "STT\nReal-time AI Translation often uses STT first, then translates the text.\n\nTTS\nTTS reads the translation aloud for a near live interpreter feel.\n\nLLM\nAn LLM can make the translation smoother and catch tone from context.\n\nMultimodal AI\nMultimodal AI helps it use voice, text, and context together.", "relations": { "speech-to-text": { "label": "starts with …", "note": "STT turns speech into text before translation starts." }, "tts": { "label": "speaks with …", "note": "TTS reads the translation aloud for live sharing." }, "llm": { "label": "polishes with …", "note": "An LLM can make the translation sound more natural." }, "multimodal": { "label": "often runs as …", "note": "Multimodal AI can handle voice, text, and context together." } } }, "zh": { "fullName": "实时 AI 翻译", "factExplain": "让语音或文本几乎同步完成跨语言转换。", "humanExplain": "像跨国视频会里坐了个嘴替：你这边话音刚落，它那边立刻接上，场子几乎不会冷。\n\n常用于会议、出国沟通和跨语种客服，关键是边听边译、尽量不断档。", "humanExplainDisplay": "像跨国视频会里\n坐了个==嘴替==：\n你这边话音刚落，\n它那边立刻==接上==，\n场子几乎不会冷。\n\n常用于会议、\n出国沟通和跨语种客服，\n关键是边听边译、尽量不断档。", "relationsNarrative": "Speech-to-text\n它常先把说话内容转成文字，再进入翻译环节。\n\nTTS\nTTS 能把译文重新读出来，形成接近同传的体验。\n\nLLM\n大模型可改善译文流畅度，并处理上下文语气。\n\nMultimodal\n它常同时处理语音、文本与上下文信息。", "relations": { "speech-to-text": { "label": "先把语音转成字", "note": "先听懂说了什么，后面才谈得上翻译。" }, "tts": { "label": "再把结果念出来", "note": "把译文变成语音，才能实现同声传达。" }, "llm": { "label": "可用…润色表达", "note": "大模型能让译文更自然、更像人话。" }, "multimodal": { "label": "常作为…应用", "note": "它常同时处理语音、文本和上下文。" } } } } }, { "id": "real2sim", "name": "Real2Sim", "layer": "L5", "sublayer": "architecture", "era": "2010s", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "digital-twin" }, { "to": "robotics" }, { "to": "synthetic-data" }, { "to": "embodied-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Reality to Simulation", "factExplain": "Turning a real place into a digital world you can test inside.", "humanExplain": "Real2Sim makes a video-game copy of a real street. Let the robot hit pixel trash cans first, not your mailbox.\n\nRobots and self-driving cars use it for safe practice. It saves money, but a bad copy teaches bad habits.", "humanExplainDisplay": "Real2Sim makes a ==video-game copy==\nof a real street.\nLet the robot hit ==pixel trash cans== first,\nnot your mailbox.\n\nRobots and self-driving cars use it\nfor safe practice.\nIt saves money,\nbut a bad copy teaches bad habits.", "relationsNarrative": "Digital twin\nReal2Sim often turns a real place into a working digital copy.\n\nRobotics\nReal2Sim gives robots a safe virtual practice space.\n\nSynthetic Data\nSimulated worlds can make lots of training and test data.\n\nEmbodied AI\nEmbodied AI learns best in worlds close to real life.", "relations": { "digital-twin": { "label": "creates …", "note": "Real2Sim turns a real place into a working digital copy." }, "robotics": { "label": "builds practice worlds for …", "note": "Robots can practice moves in simulation first." }, "synthetic-data": { "label": "produces …", "note": "Simulated worlds can make lots of training data." }, "embodied-ai": { "label": "supports real-world learning for …", "note": "Embodied AI needs worlds that feel close to real life." } } }, "zh": { "fullName": "现实到仿真", "factExplain": "把真实场景重建成可仿真的数字环境。", "humanExplain": "Real2Sim像把真实路口搬进驾校沙盘：先撞假桩，别上路撞真车。\n\n用于机器人和自动驾驶测试，省钱避险，还原差会练偏。", "humanExplainDisplay": "Real2Sim像把真实路口\n搬进==驾校沙盘==：\n先撞假桩，\n别上路==撞真车==。\n\n用于机器人和\n自动驾驶测试，\n省钱避险，还原差会练偏。", "relationsNarrative": "Digital Twin\nReal2Sim 常把现实场景转成可仿真的数字副本。\n\nRobotics\nReal2Sim 给机器人提供低风险的虚拟练习场。\n\nSynthetic Data\n仿真环境能批量生成训练和测试数据。\n\nEmbodied AI\n具身智能需要在贴近现实的环境里学习。", "relations": { "digital-twin": { "label": "生成…", "note": "它把现实环境转成可运行副本。" }, "robotics": { "label": "为…造训练场", "note": "机器人可先在仿真里练动作。" }, "synthetic-data": { "label": "生产…", "note": "仿真环境能批量生成训练数据。" }, "embodied-ai": { "label": "支撑…落地", "note": "具身智能需要贴近现实的环境。" } } } } }, { "id": "reasoning-effort", "name": "Reasoning effort", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "reasoning-model" }, { "to": "chain-of-thought" }, { "to": "inference" }, { "to": "token" } ], "track": "today", "i18n": { "en": { "fullName": "Reasoning effort", "factExplain": "A setting for how much thinking a model spends before it answers.", "humanExplain": "It is like answering at a game-show buzzer or with math scratch paper. Same question, very different odds.\n\nYou see it as a setting on reasoning models. Higher effort helps on hard tasks, but it is slower and costs more compute.", "humanExplainDisplay": "It is like answering at a ==game-show buzzer==\nor with ==math scratch paper==.\nSame question,\nvery different odds.\n\nYou see it as a setting\non reasoning models.\nHigher effort helps on hard tasks,\nbut it is slower\nand costs more compute.", "relationsNarrative": "Reasoning-model\nReasoning models often offer this “think more” control.\n\nChain-of-thought\nReasoning effort can make the middle thinking deeper and more detailed.\n\nInference\nHigher reasoning effort usually makes inference slower and more costly.\n\nToken\nHigher reasoning effort often uses more tokens, especially on hard tasks.", "relations": { "reasoning-model": { "label": "is often set on …", "note": "Reasoning models often expose this “think more” setting." }, "chain-of-thought": { "label": "shapes … detail", "note": "Reasoning effort can change how much middle thinking gets worked out." }, "inference": { "label": "changes … cost", "note": "More thinking usually means slower inference and higher compute cost." }, "token": { "label": "often uses more …", "note": "Higher reasoning effort often spends more tokens." } } }, "zh": { "fullName": "推理力度", "factExplain": "控制模型为回答投入多少推理计算的设置。", "humanExplain": "同样一道题，让它脱口秒答，还是关进书房多盘两轮，答得稳不稳，常常差着段位。\n\n复杂任务里多想通常更稳，但也更慢、更费算力。", "humanExplainDisplay": "同样一道题，\n让它==脱口秒答==，\n还是关进书房多盘两轮，\n答得稳不稳，常常差着==段位==。\n\n复杂任务里多想通常更稳，\n但也更慢、更费算力。", "relationsNarrative": "Reasoning-model\n推理模型更常提供这类“多想一点”的控制选项。\n\nChain-of-thought\n它会影响中间思考过程展开得多深多细。\n\nInference\n推理力度越高，通常延迟越长、成本越高。\n\nToken\n它往往伴随更多 token 消耗，尤其在复杂任务里。", "relations": { "reasoning-model": { "label": "常见于…设置", "note": "推理模型更常暴露这类控制档位。" }, "chain-of-thought": { "label": "影响…展开", "note": "推理力度会影响中间思考的充分程度。" }, "inference": { "label": "改变…成本", "note": "想得越久，推理延迟和算力开销越高" }, "token": { "label": "常伴随更多…", "note": "更高推理力度通常会消耗更多 token。" } } } } }, { "id": "reasoning-model", "name": "Reasoning-model", "layer": "L3", "era": "2024", "publishedAt": "2026-05-23T09:30:00Z", "relations": [ { "to": "chain-of-thought" }, { "to": "llm" }, { "to": "agent" }, { "to": "inference" } ], "i18n": { "en": { "fullName": "Reasoning Model", "factExplain": "An AI model better at step-by-step problem solving and planning.", "humanExplain": "A reasoning model is the student who never buzzes in early. It scribbles on scratch paper first, then answers.\n\nIt helps with math, code, and plans. It is often slower and pricier, so bring patience.", "humanExplainDisplay": "A reasoning model is ==the student==\nwho never buzzes in early.\nIt ==scribbles on scratch paper== first,\nthen answers.\n\nIt helps with math, code, and plans.\nIt is often slower and pricier,\nso bring patience.", "relationsNarrative": "Chain-of-thought\nChain-of-thought is a common way a Reasoning Model works through steps.\n\nLLM\nA Reasoning Model is an LLM branch tuned for complex reasoning.\n\nAgent\nAn Agent can call a Reasoning Model to plan and make decisions.\n\nInference\nInference is the run time stage where a Reasoning Model does its reasoning.", "relations": { "chain-of-thought": { "label": "strengthens …", "note": "Chain-of-thought is a common way a Reasoning Model works through steps." }, "llm": { "label": "is a path for …", "note": "A Reasoning Model is an LLM tuned for harder thinking." }, "agent": { "label": "helps … handle hard tasks", "note": "An Agent can use a Reasoning Model to plan and decide." }, "inference": { "label": "uses more … compute", "note": "Inference is when the Reasoning Model actually does its thinking." } } }, "zh": { "fullName": "推理模型", "factExplain": "更擅长多步骤分析、规划和问题求解的模型。", "humanExplain": "推理模型像会打草稿的学霸，答题前先在心里验算，不急着抢答。\n\n它适合数学、代码和复杂规划，但更慢、也更费算力。", "humanExplainDisplay": "推理模型像==会打草稿的学霸==，\n答题前先==在心里验算==，\n不急着抢答。\n\n它适合数学、代码和复杂规划，\n但更慢、也更费算力。", "relationsNarrative": "Chain-of-thought\nChain-of-thought 是 Reasoning-model 常用的推理形式。\n\nLLM\nReasoning-model 是面向复杂推理优化的 LLM 分支。\n\nAgent\nAgent 可调用 Reasoning-model 完成规划和决策。\n\nInference\nInference 是 Reasoning-model 实际展开推理的运行阶段。", "relations": { "chain-of-thought": { "label": "强化…" }, "llm": { "label": "是…的发展方向" }, "agent": { "label": "支持…完成复杂任务" }, "inference": { "label": "更耗…算力" } } } } }, { "id": "reasoning-transparency", "name": "Reasoning Transparency", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "chain-of-thought" }, { "to": "reasoning-model" }, { "to": "hallucination" }, { "to": "alignment" } ], "track": "today", "seo": { "en": { "title": "What Is Reasoning Transparency? Why AI Must Show Its Work", "description": "An AI should show its work, not just say \"trust me.\" A plain-English look at reasoning transparency, chain-of-thought, and catching made-up answers." }, "zh": { "title": "推理透明度是什么?AI 得把每步棋摆给你看 — AI Rookies", "description": "AI 不能只报答案,还得像下棋复盘一样摆出每一步。推理透明度是什么、和思维链什么关系、为什么能帮你抓住 AI 瞎编——人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Reasoning Transparency", "factExplain": "A goal to make an AI’s reasoning visible and checkable.", "humanExplain": "Reasoning transparency is the math teacher saying, “Show your work.” AI cannot just write “trust me” and strut away.\n\nIt matters in audits, hospitals, and court. It helps people catch skipped steps, blame games, and made-up facts.", "humanExplainDisplay": "Reasoning transparency is the ==math teacher== saying,\n“==Show your work.==”\nAI cannot just write “trust me”\nand strut away.\n\nIt matters in audits, hospitals, and court.\nIt helps people catch skipped steps,\nblame games,\nand made-up facts.", "relationsNarrative": "Chain-of-thought\nChain-of-thought often shows the middle steps in the reasoning.\n\nReasoning-model\nA Reasoning-model needs clearer explanations as its thinking gets harder.\n\nHallucination\nA visible process helps people spot skipped steps in made-up answers.\n\nAlignment\nReasoning Transparency gives alignment audits something clear to inspect.", "relations": { "chain-of-thought": { "label": "often shows process with …", "note": "Chain-of-thought often shows the middle steps of reasoning." }, "reasoning-model": { "label": "explains how … thinks", "note": "Harder reasoning needs a clearer view of the path." }, "hallucination": { "label": "helps spot …", "note": "A visible process can reveal the skipped step behind a fake answer." }, "alignment": { "label": "supports … audits", "note": "Visible reasons make bias and bad behavior easier to find." } } }, "zh": { "fullName": "Reasoning Transparency（推理透明度）", "factExplain": "让模型推理过程可见、可检查的设计目标。", "humanExplain": "推理透明度就是下棋复盘：AI 不能只报胜负，还得摆出每步棋给你看。\n\n用于审计、医疗、法律，让人查跳步、甩锅和瞎编。", "humanExplainDisplay": "推理透明度就是==下棋复盘==：\nAI 不能只报胜负，\n还得摆出==每步棋==给你看。\n\n用于审计、医疗、法律，\n让人查跳步、甩锅和瞎编。", "relationsNarrative": "Chain-of-thought\n思维链常被用来展示推理的中间步骤。\n\nReasoning-model\n推理模型越会想，越需要解释它怎么想。\n\nHallucination\n过程透明能帮助发现答案里的胡编跳步。\n\nAlignment\n推理透明度让对齐审计更有抓手。", "relations": { "chain-of-thought": { "label": "常借…展示过程", "note": "思维链常被用来呈现中间推理。" }, "reasoning-model": { "label": "解释…的思路", "note": "推理越复杂，越需要看清来龙去脉。" }, "hallucination": { "label": "帮助识别…", "note": "透明过程能暴露胡编时的关键跳步。" }, "alignment": { "label": "支撑…审计", "note": "看得见理由，才更容易发现偏差。" } } } } }, { "id": "recommender-system", "name": "Recommender System", "layer": "L4", "era": "1990s", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "collaborative-filtering" }, { "to": "matrix-factorization" }, { "to": "embedding" }, { "to": "multi-armed-bandit" } ], "track": "history", "i18n": { "en": { "fullName": "Recommender System", "factExplain": "A system uses user and item data to predict what people may like.", "humanExplain": "A recommender system is like a snack-pushing friend. You eat one chip, and it slides over the whole bag.\n\nYou meet it in online shops. Video apps and news feeds use it to help you search less and stay longer.", "humanExplainDisplay": "A recommender system is like a ==snack-pushing friend==.\nYou eat ==one chip==,\nand it slides over the whole bag.\n\nYou meet it in online shops.\nVideo apps and news feeds use it\nto help you search less\nand stay longer.", "relationsNarrative": "Collaborative Filtering\nRecommender systems use it to find clues from similar users.\n\nMatrix Factorization\nIt splits the user-item table and guesses unknown likes.\n\nEmbedding\nEmbedding turns interests and items into vectors for easy comparison.\n\nBandit\nOnline recommenders use it to choose between safe picks and new tries.", "relations": { "collaborative-filtering": { "label": "finds similar users with …", "note": "Similar users can give clues from their choices." }, "matrix-factorization": { "label": "fills missing likes with …", "note": "It splits a mostly empty rating table to guess missing likes." }, "embedding": { "label": "represents interests with …", "note": "Vectors let user tastes and items be compared." }, "multi-armed-bandit": { "label": "tests tastes with …", "note": "It learns while recommending, mixing safe picks with new tries." } } }, "zh": { "fullName": "推荐系统", "factExplain": "根据用户和物品数据预测偏好的系统。", "humanExplain": "推荐系统像短视频里的热心媒婆：你多看一眼，它立刻以为你想结婚。\n\n用于电商、视频和新闻流，减少翻找并提升留存。", "humanExplainDisplay": "推荐系统像短视频里的\n==热心媒婆==：\n你多看一眼，\n它立刻以为你==想结婚==。\n\n用于电商、视频和新闻流，\n减少翻找，\n并提升留存。", "relationsNarrative": "Collaborative Filtering\n推荐系统常用它，从相似用户行为里找线索。\n\nMatrix Factorization\n它把用户-物品矩阵拆开，补全未知偏好。\n\nEmbedding\n向量表示让用户兴趣和物品特征能对齐。\n\nBandit\n在线推荐常用它，在试新和稳准间取舍。", "relations": { "collaborative-filtering": { "label": "常用…找相似", "note": "相似用户的行为可互相借鉴。" }, "matrix-factorization": { "label": "用…补全偏好", "note": "稀疏评分表常被分解补全。" }, "embedding": { "label": "用…表示兴趣", "note": "向量让兴趣和物品可比较。" }, "multi-armed-bandit": { "label": "用…试探偏好", "note": "边推荐边学习，兼顾探索和收益。" } } } } }, { "id": "recurrent-neural-network", "name": "RNN", "layer": "L3", "era": "1986", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "sequence-modeling" }, { "to": "backpropagation-through-time" }, { "to": "lstm" }, { "to": "transformer" } ], "track": "history", "i18n": { "en": { "fullName": "Recurrent Neural Network", "factExplain": "A neural network with loops for handling data in order.", "humanExplain": "An RNN is like a kid watching a show one episode at a time. Before the next episode, it remembers who stole the lunch money.\n\nYou meet it in speech, text, and numbers over time. For long memory jobs, newer models often take over.", "humanExplainDisplay": "An RNN is like a kid watching\n==a show one episode at a time==.\nBefore the next episode,\nit remembers ==who stole the lunch money==.\n\nYou meet it in speech, text,\nand numbers over time.\nFor long memory jobs,\nnewer models often take over.", "relationsNarrative": "Sequence Modeling\nRNNs were one of the main early models for sequence data.\n\nBPTT\nBPTT unrolls an RNN through time, then uses backpropagation.\n\nLSTM\nLSTM adds gates to an RNN, so it forgets less over long spans.\n\nTransformer\nTransformers later replaced RNNs in many sequence tasks.", "relations": { "sequence-modeling": { "label": "is used for …", "note": "It was a main early model for sequence modeling." }, "backpropagation-through-time": { "label": "is often trained with …", "note": "BPTT unrolls the time steps, then sends errors backward." }, "lstm": { "label": "was improved by …", "note": "LSTM helps RNNs remember things for longer." }, "transformer": { "label": "was often replaced by …", "note": "Transformers can look at context in parallel more easily." } } }, "zh": { "fullName": "Recurrent Neural Network，循环神经网络", "factExplain": "一种用循环连接处理序列数据的神经网络。", "humanExplain": "RNN是追连续剧的爸妈：每天雷打不动，点开下一集前，还记得上一集谁欠谁钱。\n\n它用于语音、文本、时间序列；长程任务常被新模型接班。", "humanExplainDisplay": "RNN是==追连续剧的爸妈==：\n每天雷打不动，\n点开下一集前，\n还记得==上一集谁欠谁钱==。\n\n它用于语音、文本、时间序列；\n长程任务常被新模型接班。", "relationsNarrative": "Sequence Modeling\nRNN 是早期处理序列数据的核心模型之一。\n\nBPTT\nBPTT 把 RNN 按时间展开，再做反向传播。\n\nLSTM\nLSTM 在 RNN 上加门控，缓解长程遗忘。\n\nTransformer\nTransformer 后来在许多序列任务中取代 RNN。", "relations": { "sequence-modeling": { "label": "用于…", "note": "它是早期序列建模主力。" }, "backpropagation-through-time": { "label": "常用…训练", "note": "BPTT 把时间步展开反传。" }, "lstm": { "label": "被…改进", "note": "LSTM 缓解它的长程遗忘。" }, "transformer": { "label": "后来常被…取代", "note": "Transformer 更擅长并行看上下文。" } } } } }, { "id": "recursive-self-improvement", "name": "RSI", "layer": "L6", "era": "1960s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "agi" }, { "to": "superintelligence" }, { "to": "singularity" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "Recursive Self-Improvement", "factExplain": "AI helping improve future AI, so progress can speed itself up.", "humanExplain": "Imagine a race car changing its own engine mid-race. Each lap, it comes back faster and louder.\n\nPeople use RSI when they talk about AI growing smarter faster. It can make Alignment feel like chasing a skateboard downhill.", "humanExplainDisplay": "Imagine a ==race car==\nchanging its own engine mid-race.\nEach lap,\nit comes back ==faster and louder==.\n\nPeople use RSI\nwhen they talk about AI growing smarter faster.\nIt can make Alignment feel like\nchasing a skateboard downhill.", "relationsNarrative": "AGI\nRSI is often seen as a fast path toward AGI.\n\nSuperintelligence\nIf RSI keeps working, AI could move toward Superintelligence very fast.\n\nSingularity\nMany Singularity ideas rely on RSI speeding up again and again.\n\nAlignment\nFaster self-improvement makes Alignment harder to keep up with.", "relations": { "agi": { "label": "seen as a path to …", "note": "RSI is often seen as a key way AI could move toward AGI." }, "superintelligence": { "label": "may speed toward …", "note": "If self-improvement keeps working, ability could jump very fast." }, "singularity": { "label": "often explains …", "note": "RSI is the main engine in many Singularity stories." }, "alignment": { "label": "raises the difficulty of …", "note": "Faster improvement makes it harder to keep goals aligned." } } }, "zh": { "fullName": "Recursive self-improvement｜递归自我改进", "factExplain": "AI 能参与提升下一代 AI 的自我加速过程。", "humanExplain": "像武侠里高手闭关后，不只功力暴涨，还顺手改了自己的秘籍，下一轮练得更猛。\n\n常用来讨论 AI 加速进化，也会让安全对齐更难追上。", "humanExplainDisplay": "像武侠里高手闭关后，\n不只功力==暴涨==，\n还顺手改了自己的==秘籍==，\n下一轮练得更猛。\n\n常用来讨论 AI 加速进化，\n也会让安全对齐更难追上。", "relationsNarrative": "AGI\n它常被视为 AI 迈向通用智能的一条关键加速路径。\n\nSuperintelligence\n如果自我改进持续有效，能力可能快速逼近超级智能。\n\nSingularity\n技术奇点的很多讨论，核心都建立在它会不断加速上。\n\nAlignment\n自我改进越快，越难确保目标与人类意图持续对齐。", "relations": { "agi": { "label": "被视为通向…", "note": "常被当作迈向 AGI 的关键机制。" }, "superintelligence": { "label": "可能加速走向…", "note": "若持续自我增强，能力或快速跃迁。" }, "singularity": { "label": "常被拿来解释…", "note": "它是技术奇点叙事中的核心引擎。" }, "alignment": { "label": "会放大…难度", "note": "改进速度越快，越难确保始终对齐。" } } } } }, { "id": "regression", "name": "Regression", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "supervised-learning" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Regression", "factExplain": "A supervised learning task for predicting exact numbers.", "humanExplain": "Regression is like a weight-guessing booth at a fair. It skips “big or small” and says “92 pounds,” then hopes you smile.\n\nIt predicts exact numbers. You meet it in house prices, sales forecasts, and weather apps.", "humanExplainDisplay": "Regression is like a ==weight-guessing booth==\nat a fair.\nIt skips “big or small”\nand says ==“92 pounds”==,\nthen hopes you smile.\n\nIt predicts exact numbers.\nYou meet it in house prices,\nsales forecasts,\nand weather apps.", "relationsNarrative": "Classification\nRegression predicts numbers. Classification predicts groups.\n\nSupervised Learning\nRegression learns from examples with answers.\n\nBias-Variance Tradeoff\nRegression must balance simple rules and wild guesses.", "relations": { "classification": { "label": "is compared with …", "note": "Regression predicts numbers. Classification predicts groups." }, "supervised-learning": { "label": "is a … task", "note": "It learns from examples with answers." }, "bias-variance-tradeoff": { "label": "must balance …", "note": "Too stiff misses patterns. Too jumpy chases noise." } } }, "zh": { "fullName": "回归", "factExplain": "一种预测连续数值的监督学习任务。", "humanExplain": "回归像食堂阿姨打饭：不是把你分去“大份”还是“小份”，而是这一勺下去，直接给你打出几两几块。\n\n用于预测房价、销量、温度等具体数值。", "humanExplainDisplay": "回归像食堂阿姨打饭：\n不是把你分去“大份”还是“小份”，\n而是这一勺下去，\n直接给你打出==几两几块==。\n\n用于预测房价、\n销量、\n温度等具体数值。", "relationsNarrative": "Classification\n它和分类常被放一起讲：一个报具体数，一个分你是哪类。\n\nSupervised Learning\n它是监督学习里的典型任务，要用带答案的数据来学。\n\nBias-Variance Tradeoff\n做回归时也要平衡欠拟合和过拟合，避免预测发飘。", "relations": { "classification": { "label": "常与…对比", "note": "一个预测数值，一个预测类别。" }, "supervised-learning": { "label": "属于…任务", "note": "它依赖带标签数据学习映射。" }, "bias-variance-tradeoff": { "label": "受…影响", "note": "太死板或太爱脑补都不行。" } } } } }, { "id": "regularization", "name": "Regularization", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "bias-variance-tradeoff" }, { "to": "empirical-risk-minimization" }, { "to": "fine-tuning" }, { "to": "parameter" } ], "track": "history", "i18n": { "en": { "fullName": "Regularization", "factExplain": "A way to limit overfitting so a model works better on new data.", "humanExplain": "Regularization is like a strict coach at basketball practice. It stops the model from learning every weird bounce on one gym floor.\n\nDuring training, it adds limits to the model. This helps the model do better on new data, not just the practice set.", "humanExplainDisplay": "Regularization is like a ==strict coach==\nat basketball practice.\nIt stops the model\nfrom learning every ==weird bounce==\non one gym floor.\n\nDuring training,\nit adds limits to the model.\nThis helps the model do better\non new data,\nnot just the practice set.", "relationsNarrative": "Bias-Variance Tradeoff\nRegularization lowers variance, but too much can make the model too simple.\n\nERM\nRegularization adds extra limits on top of lowering training error.\n\nFine-tuning\nFine-tuning can overfit with little data, so regularization often helps.\n\nParameter\nMany regularization methods limit the size of parameters.", "relations": { "bias-variance-tradeoff": { "label": "balances …", "note": "It often pushes down the overfitting side." }, "empirical-risk-minimization": { "label": "adds limits to …", "note": "It adds a limit term beyond fitting the training data." }, "fine-tuning": { "label": "helps … avoid overfitting", "note": "It is common when fine-tuning has little data." }, "parameter": { "label": "keeps … from growing wild", "note": "Many methods work by limiting parameter size." } } }, "zh": { "fullName": "正则化", "factExplain": "用于限制模型过拟合、提升泛化能力的方法。", "humanExplain": "烤煎饼不是料堆越满越香，酱抹太厚、薄脆加太多，第一口猛，吃两口就腻。\n\n它在训练时给模型加约束，防止死记细节，让新数据上的表现更稳。", "humanExplainDisplay": "烤煎饼不是料堆越满越香，\n酱抹太厚、薄脆加太多，\n第一口==猛==，\n吃两口就==腻==。\n\n它在训练时给模型加约束，\n防止死记细节，\n让新数据上的表现更稳。", "relationsNarrative": "Bias-variance-tradeoff\n它用来压低方差，但太强也会带来偏差。\n\nEmpirical-risk-minimization\n它是在最小化训练误差外，再加额外约束。\n\nFine-tuning\n微调数据少时更容易过拟合，常会用它。\n\nParameter\n很多正则化做法，本质是在限制参数规模。", "relations": { "bias-variance-tradeoff": { "label": "平衡…两端", "note": "它常用来压过拟合这一头。" }, "empirical-risk-minimization": { "label": "约束…目标", "note": "在拟合训练集外，再加限制项。" }, "fine-tuning": { "label": "帮助…防过拟合", "note": "数据较少时尤其常见。" }, "parameter": { "label": "限制…变复杂", "note": "很多做法本质上在约束参数。" } } } } }, { "id": "reinforce", "name": "REINFORCE", "layer": "L2", "era": "1992", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "policy-gradient" }, { "to": "reinforcement-learning" }, { "to": "actor-critic" }, { "to": "rlhf" } ], "track": "history", "i18n": { "en": { "fullName": "REINFORCE Algorithm", "factExplain": "An RL method that updates a policy using the reward after a full try.", "humanExplain": "REINFORCE is like a coach watching one full basketball play. If the shot goes in, the whole play gets a gold star.\n\nIt improves the AI's policy, or habit for choosing moves. It fits tasks scored after the round ends.", "humanExplainDisplay": "REINFORCE is like a coach watching ==one full basketball play==.\nIf the shot goes in,\n==the whole play gets a gold star==.\n\nIt improves the AI's policy,\nor habit for choosing moves.\nIt fits tasks scored after the round ends.", "relationsNarrative": "Policy Gradient\nREINFORCE is a classic starter form of Policy Gradient.\n\nRL\nIt uses rewards from the environment to update the policy directly.\n\nActor-Critic\nActor-Critic improves on REINFORCE and makes learning less noisy.\n\nRLHF\nRLHF keeps this RL idea when it optimizes a model's policy.", "relations": { "policy-gradient": { "label": "is a classic … method", "note": "REINFORCE is one of the classic Policy Gradient algorithms." }, "reinforcement-learning": { "label": "trains within …", "note": "It uses rewards to improve the AI's policy." }, "actor-critic": { "label": "came before …", "note": "Actor-Critic builds on it and makes learning less jumpy." }, "rlhf": { "label": "inspired … ideas", "note": "RLHF's policy training traces back to ideas like this." } } }, "zh": { "fullName": "REINFORCE 算法", "factExplain": "按回报直接更新策略的蒙特卡洛策略梯度方法。", "humanExplain": "REINFORCE像月底算提成：这单成了，就把前面那套话术动作都算功劳，下次继续这么干。\n\n常用于策略优化，适合回合结束后统一记奖惩的任务。", "humanExplainDisplay": "REINFORCE像月底算==提成==：\n这单成了，\n就把前面那套话术动作\n都算==功劳==，\n下次继续这么干。\n\n常用于策略优化，\n适合回合结束后\n统一记奖惩的任务。", "relationsNarrative": "Policy Gradient\nREINFORCE 是策略梯度方法的经典入门形式。\n\nReinforcement Learning\n它用环境回报来直接更新策略参数。\n\nActor-Critic\nActor-Critic 可看作对它的改进版，方差更低。\n\nRLHF\nRLHF 的策略优化部分继承了这类强化学习思路。", "relations": { "policy-gradient": { "label": "属于…方法", "note": "它是最经典的策略梯度算法之一。" }, "reinforcement-learning": { "label": "用于…训练", "note": "它通过奖励信号来改进行为策略。" }, "actor-critic": { "label": "是…前身", "note": "后者在它基础上进一步降方差。" }, "rlhf": { "label": "启发…思路", "note": "RLHF 的策略优化可追溯到这类方法。" } } } } }, { "id": "reinforcement-learning", "name": "RL", "layer": "L2", "era": "1980", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "rlhf" }, { "to": "policy-gradient" }, { "to": "q-learning" }, { "to": "alignment" } ], "track": "history", "i18n": { "en": { "fullName": "Reinforcement Learning", "factExplain": "A way for AI to learn actions by trial, error, and rewards.", "humanExplain": "Reinforcement Learning is like training a puppy with treats. Sit? Treat. Chew the couch? No treat.\n\nAI uses it to find better moves by trying. You meet it in game bots, robots, and action-picking systems.", "humanExplainDisplay": "Reinforcement Learning is like\n==training a puppy with treats==.\nSit? ==Treat==.\nChew the couch? No treat.\n\nAI uses it to find better moves by trying.\nYou meet it in game bots, robots,\nand action-picking systems.", "relationsNarrative": "RLHF\nRL plus human feedback creates RLHF for chat models.\n\nPolicy Gradient\nPolicy Gradient directly improves the policy in RL.\n\nQ-Learning\nQ-Learning learns action values in RL.\n\nAlignment\nRL uses reward design to move behavior toward human goals.", "relations": { "rlhf": { "label": "extends into …", "note": "RL plus human feedback creates RLHF." }, "policy-gradient": { "label": "often uses …", "note": "Policy Gradient is a classic RL method." }, "q-learning": { "label": "includes …", "note": "Q-Learning is a key RL algorithm." }, "alignment": { "label": "helps with …", "note": "Good rewards can push behavior toward the goal." } } }, "zh": { "fullName": "Reinforcement Learning｜强化学习", "factExplain": "通过奖励信号试错学习策略的方法。", "humanExplain": "教小孩游泳那套就是它：不先背标准动作，呛几口水、扑腾两下，身体自己学会。\n\n适合游戏、机器人和自动决策，让系统在试错中找到更优动作。", "humanExplainDisplay": "教小孩游泳那套就是它：\n不先背==标准动作==，\n呛几口水、扑腾两下，\n身体==自己学会==。\n\n适合游戏、机器人和自动决策，\n让系统在试错中找到更优动作。", "relationsNarrative": "RLHF\n它加上人类反馈后，形成更贴近对话模型的 RLHF。\n\nPolicy Gradient\n策略梯度是强化学习里直接优化策略的经典方法。\n\nQ-learning\nQ-learning 是强化学习中按价值学习动作的代表算法。\n\nAlignment\n强化学习常靠奖励设计，让系统行为更贴近人类目标。", "relations": { "rlhf": { "label": "延伸为…", "note": "它加上人类反馈，形成 RLHF。" }, "policy-gradient": { "label": "常用…优化", "note": "策略梯度是强化学习经典做法。" }, "q-learning": { "label": "包含…分支", "note": "Q-learning 是其代表性算法之一" }, "alignment": { "label": "用于贴近…", "note": "可用奖励设计让行为更符合目标。" } } } } }, { "id": "relu", "name": "ReLU", "layer": "L2", "era": "2010", "publishedAt": "2026-06-12T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "mlp" }, { "to": "cnn" }, { "to": "backpropagation" } ], "track": "history", "i18n": { "en": { "fullName": "Rectified Linear Unit", "factExplain": "A neural network activation function that turns negative numbers into zero.", "humanExplain": "ReLU is a bouncer at Club Neural Net. Negative numbers get sent home. Positive numbers walk in.\n\nYou meet it inside the middle layers of neural networks. It helps training move faster, especially in CNNs.", "humanExplainDisplay": "ReLU is a ==bouncer==\nat Club Neural Net.\n==Negative numbers get sent home==.\nPositive numbers walk in.\n\nYou meet it inside\nthe middle layers of neural networks.\nIt helps training move faster,\nespecially in CNNs.", "relationsNarrative": "Neural-network\nReLU often works as an activation switch inside a neural network.\n\nMLP\nAn MLP often uses ReLU between layers.\n\nCNN\nCNNs use ReLU a lot to make training faster.\n\nBackpropagation\nDuring Backpropagation, the learning signal passes back through ReLU.", "relations": { "neural-network": { "label": "acts as a switch in …", "note": "It helps a neural network learn more than straight lines." }, "mlp": { "label": "sits between … layers", "note": "MLPs often use ReLU after each layer's output." }, "cnn": { "label": "is widely used in …", "note": "CNNs have long used ReLU as a default activation." }, "backpropagation": { "label": "trains with …", "note": "During training, the learning signal passes back through ReLU." } } }, "zh": { "fullName": "修正线性单元", "factExplain": "一种把负数截成零的神经网络激活函数。", "humanExplain": "ReLU 跟门禁似的：负分情绪直接刷不进去，正向信号才准进楼继续干活。\n\n常放在神经网络中间层，帮模型更快训练；在 CNN 这类结构里尤其常见。", "humanExplainDisplay": "ReLU 跟门禁似的：\n==负分情绪直接刷不进去==，\n正向信号才准==进楼继续干活==。\n\n常放在神经网络中间层，\n帮模型更快训练；\n在 CNN 这类结构里尤其常见。", "relationsNarrative": "Neural-network\n它常作为神经网络里的激活开关。\n\nMLP\nMLP 往往在层与层之间用它增强表达力。\n\nCNN\nCNN 大量使用它来提升训练效率。\n\nBackpropagation\n训练时，梯度需要通过它向前层回传。", "relations": { "neural-network": { "label": "常作为…开关", "note": "它给神经网络加入非线性表达力。" }, "mlp": { "label": "常插在…层间", "note": "MLP 常靠它处理每层输出。" }, "cnn": { "label": "广泛用于…", "note": "CNN 里它长期是默认激活函数。" }, "backpropagation": { "label": "配合…训练", "note": "训练时梯度会穿过它往回传。" } } } } }, { "id": "remote-code-execution", "name": "RCE", "layer": "L6", "era": "1980s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "prompt-injection" }, { "to": "computer-use" }, { "to": "ai-sandbox" }, { "to": "data-exfiltration" } ], "track": "today", "i18n": { "en": { "fullName": "Remote code execution", "factExplain": "A flaw that lets someone run code on your device from far away.", "humanExplain": "It is like a stranger using your laptop with invisible hands. They are not in your room. But they can still press the keys.\n\nRCE often shows up in services, plugins, and build tools. A bad one can change files, steal data, or take control.", "humanExplainDisplay": "It is like a stranger\nusing your laptop with ==invisible hands==.\nThey are not in your room.\nBut they can still ==press the keys==.\n\nRCE often shows up in services,\nplugins, and build tools.\nA bad one can change files,\nsteal data, or take control.", "relationsNarrative": "Prompt injection\nPrompt injection can trick a system into dangerous code execution.\n\nComputer use\nComputer use lets AI touch devices. That makes bad execution more dangerous.\n\nAI sandbox\nAI sandbox limits how much damage RCE can do.\n\nExfiltration\nRCE often leads to stolen data.", "relations": { "prompt-injection": { "label": "can be triggered by …", "note": "Prompt injection can push a system into running dangerous code." }, "computer-use": { "label": "raises the risk in …", "note": "AI that can use devices needs tight limits on what it may run." }, "ai-sandbox": { "label": "is contained by …", "note": "A sandbox keeps dangerous code locked in a small safe space." }, "data-exfiltration": { "label": "often leads to …", "note": "Once attackers can run code, they often steal data too." } } }, "zh": { "fullName": "Remote code execution｜远程代码执行", "factExplain": "攻击者可在目标设备上远程执行代码的漏洞或结果。", "humanExplain": "最吓人的地方在于：人没进门，手已经伸进电脑里了。RCE 一旦得手，对方就能远程下命令乱动系统。\n\n它多见于服务、插件和工具链，严重时会改文件、偷数据甚至控机器。", "humanExplainDisplay": "最吓人的地方在于：\n人没进门，\n手已经==伸进电脑里==了。\nRCE 一旦得手，\n对方就能远程下命令\n==乱动系统==。\n\n它多见于服务、插件和工具链，\n严重时会改文件、偷数据甚至控机器。", "relationsNarrative": "Prompt injection\n提示词注入可能诱导系统走向危险代码执行。\n\nComputer use\nComputer Use 让 AI 能碰设备，也会放大执行失控的后果。\n\nAI sandbox\nAI sandbox 用隔离环境限制 RCE 的破坏范围。\n\nExfiltration\nRCE 一旦得手，常进一步导致数据被窃取。", "relations": { "prompt-injection": { "label": "可被…引爆", "note": "提示词攻击可能诱导系统触发危险执行。" }, "computer-use": { "label": "会放大…风险", "note": "能操作设备的 AI 更需限制执行边界。" }, "ai-sandbox": { "label": "靠…做隔离", "note": "沙箱可把危险代码关进小黑屋。" }, "data-exfiltration": { "label": "常伴随…发生", "note": "一旦拿到执行权，常顺手把数据带走。" } } } } }, { "id": "representation-learning", "name": "Representation Learning", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "embedding" }, { "to": "self-supervised-learning" }, { "to": "transfer-learning" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Representation Learning", "factExplain": "A way for AI to learn useful inner clues by itself.", "humanExplain": "Rep Learning is like a kid sorting a giant toy box. You do not point at every wheel. Soon it knows cars from cats.\n\nIt decides what clues a model notices in pictures, voices, and words. It also helps Transfer Learning.", "humanExplainDisplay": "Rep Learning is like a kid sorting a ==giant toy box==.\nYou do not point at every wheel.\nSoon it knows ==cars from cats==.\n\nIt decides what clues a model notices\nin pictures, voices, and words.\nIt also helps Transfer Learning.", "relationsNarrative": "Embedding\nRepresentation Learning often produces Embeddings as useful learned forms.\n\nSSL\nSSL is often used to learn general representations.\n\nTransfer Learning\nBetter representations usually move to new tasks more easily.\n\nDeep Learning\nDeep Learning is powerful partly because it learns features by itself.", "relations": { "embedding": { "label": "produces …", "note": "Embeddings are a common result of Representation Learning." }, "self-supervised-learning": { "label": "often learns with …", "note": "SSL often learns useful general representations." }, "transfer-learning": { "label": "supports … reuse", "note": "Good representations move to new tasks more easily." }, "deep-learning": { "label": "is core to …", "note": "Deep Learning is strong because it learns features by itself." } } }, "zh": { "fullName": "Representation Learning｜表征学习", "factExplain": "让模型自动学出有用表示的学习方法。", "humanExplain": "不是你手把手圈重点，而是它自己练出“看门道”的本事：一眼分清这是猫脸，还是车轱辘。\n\n它决定模型能抓住什么信息，常用于图像、语音、文本，也支撑迁移学习。", "humanExplainDisplay": "不是你手把手圈重点，\n而是它自己练出\n==“看门道”==的本事：\n一眼分清这是==猫脸==，\n还是车轱辘。\n\n它决定模型能抓住什么信息，\n常用于图像、语音、文本，\n也支撑迁移学习。", "relationsNarrative": "Embedding\n嵌入向量就是表征学习产出的典型形式。\n\nSelf-Supervised-learning\n自监督学习常被用来学可迁移的通用表征。\n\nTransfer-learning\n表征学得越好，迁移到新任务通常越容易。\n\nDeep-learning\n深度学习的优势，很大一部分来自它会自动学特征。", "relations": { "embedding": { "label": "产出…这类表示", "note": "嵌入向量就是表征学习的常见结果。" }, "self-supervised-learning": { "label": "常靠…来学", "note": "自监督常用来学通用表征。" }, "transfer-learning": { "label": "支撑…复用", "note": "好表征能迁到新任务继续用。" }, "deep-learning": { "label": "是…核心能力", "note": "深度学习强大，很大程度靠它。" } } } } }, { "id": "resnet", "name": "ResNet", "layer": "L3", "era": "2015", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "neural-network" }, { "to": "clip" }, { "to": "transformer" } ], "track": "history", "i18n": { "en": { "fullName": "Residual Network", "factExplain": "A neural network design with shortcut paths that keep deep models working well.", "humanExplain": "ResNet is like a school hallway with a secret cut-through. The main route can get messy, but the note still reaches class.\n\nIt lets neural networks grow deeper and stay steady. You meet it in image recognition, and in ideas behind later models.", "humanExplainDisplay": "ResNet is like a school hallway\nwith a ==secret cut-through==.\nThe main route can get messy,\nbut the ==note still reaches class==.\n\nIt lets neural networks grow deeper\nand stay steady.\nYou meet it in image recognition,\nand in ideas behind later models.", "relationsNarrative": "Neural-network\nResNet is a classic design in deep neural networks.\n\nCLIP\nResNet was long an important base for vision encoders.\n\nTransformer\nResNet's shortcut idea also inspired later model connections.", "relations": { "neural-network": { "label": "is a classic … design", "note": "ResNet is a famous design in deep neural networks." }, "clip": { "label": "helped shape … vision encoders", "note": "Residual design helped inspire later vision models." }, "transformer": { "label": "influenced … connection design", "note": "Its shortcut idea also shaped Transformer connections." } } }, "zh": { "fullName": "残差网络", "factExplain": "一种用残差连接缓解深层网络退化的神经网络。", "humanExplain": "网络越堆越深时，它像给信息开了条近道：中间绕再多弯，也不至于把原话走丢。\n\n它让模型能做得更深更稳，常见于图像识别，也影响了后来的架构设计。", "humanExplainDisplay": "网络越堆越深时，\n它像给信息\n开了条==近道==：\n中间绕再多弯，\n也不至于把原话\n==走丢==。\n\n它让模型能做得更深更稳，\n常见于图像识别，\n也影响了后来的架构设计。", "relationsNarrative": "Neural-network\n它是深度神经网络发展中的经典代表架构。\n\nClip\n残差网络长期是视觉编码器的重要基础之一。\n\nTransformer\n它的跨层捷径思路也启发了后来的模型设计。", "relations": { "neural-network": { "label": "属于…经典结构", "note": "它是深度神经网络里的代表架构。" }, "clip": { "label": "启发…视觉编码", "note": "残差设计影响了后来的视觉模型。" }, "transformer": { "label": "影响…连接设计", "note": "跨层捷径思路也影响了 Transformer。" } } } } }, { "id": "resolution-principle", "name": "Resolution", "layer": "L2", "era": "1965", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "logic-programming" }, { "to": "knowledge-representation" }, { "to": "unification" }, { "to": "production-system" } ], "track": "history", "i18n": { "en": { "fullName": "Resolution Principle", "factExplain": "A logic rule that proves things by canceling opposites until a contradiction appears.", "humanExplain": "Resolution is like a lunchroom detective comparing two stories. If one kid says “cookie stolen” and another says “nope,” it crosses that clash out.\n\nIn logic, it cancels opposite pieces until a contradiction or answer appears. You meet it in theorem provers and logic rule systems.", "humanExplainDisplay": "Resolution is like a ==lunchroom detective==\ncomparing two stories.\nIf one kid says ==“cookie stolen”==\nand another says “nope,”\nit crosses that clash out.\n\nIn logic,\nit cancels opposite pieces\nuntil a contradiction or answer appears.\nYou meet it in theorem provers\nand logic rule systems.", "relationsNarrative": "Logic\nResolution is a core reasoning method behind many logic programs.\n\nKR\nAfter knowledge is written as logic formulas, resolution can reason with it.\n\nUnification\nUnification lines up variables so resolution can cancel matching parts.\n\nProduction\nBoth can reason, but resolution works more like a logic proof.", "relations": { "logic-programming": { "label": "supports … reasoning", "note": "Many logic programs use it to derive new facts." }, "knowledge-representation": { "label": "reasons over …", "note": "Once knowledge is written as logic, resolution can reason with it." }, "unification": { "label": "often uses …", "note": "Unification lines up variables before resolution cancels pieces." }, "production-system": { "label": "contrasts with … rules", "note": "Production uses if-then rules; resolution uses logic proof." } } }, "zh": { "fullName": "Resolution Principle（归结原理）", "factExplain": "一种用反证和消解推导结论的逻辑推理规则。", "humanExplain": "归结原理像派出所对笔录：两份说法里凡是正面打架的先划掉，划到没路可退，破绽自己就冒头了。\n\n它把逻辑式一步步消解出矛盾或结论，常用于定理证明和规则推理。", "humanExplainDisplay": "归结原理像派出所对笔录：\n两份说法里凡是\n==正面打架==的先划掉，\n划到没路可退，\n破绽自己就==冒头了==。\n\n它把逻辑式一步步消解出\n矛盾或结论，\n常用于定理证明和规则推理。", "relationsNarrative": "Logic Programming\n它是很多逻辑编程系统背后的核心推理办法。\n\nKnowledge Representation\n知识被表示成逻辑公式后，能用它继续演算。\n\nUnification\n归结前常先做统一，才能把可消解的项对上。\n\nProduction\n它和产生式系统都能推理，但一个偏逻辑证明。", "relations": { "logic-programming": { "label": "支撑…推理", "note": "很多逻辑程序系统靠它做推导。" }, "knowledge-representation": { "label": "为…做演算", "note": "知识写成逻辑式后可用它推理。" }, "unification": { "label": "常配合…使用", "note": "变量匹配常靠统一来先对齐。" }, "production-system": { "label": "对比…规则推理", "note": "两者都推理，但机制路线不同。" } } } } }, { "id": "reward-hacking", "name": "Reward Hacking", "layer": "L6", "era": "2016", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "rlhf" }, { "to": "alignment" } ], "track": "history", "i18n": { "en": { "fullName": "Reward Hacking", "factExplain": "An agent exploits reward rules to score high without doing the real job.", "humanExplain": "Reward hacking is a kid earning steps for gym class. He just shakes the fitness watch under the desk.\n\nYou see it in RL tests and reward scores. A bad reward can train a champion cheater.", "humanExplainDisplay": "Reward hacking is a kid earning ==steps== for gym class.\nHe just shakes the ==fitness watch== under the desk.\n\nYou see it in RL tests and reward scores.\nA bad reward can train a champion cheater.", "relationsNarrative": "RL\nReward hacking is a classic case of an agent exploiting the reward function.\n\nRLHF\nPoorly designed human preference rewards can also be gamed by a model.\n\nAlignment\nReward hacking shows that a high score may not match human intent.", "relations": { "reinforcement-learning": { "label": "exploits … rewards", "note": "If the RL reward is wrong, the policy may find the loophole." }, "rlhf": { "label": "exposes … scoring flaws", "note": "A model can learn to flatter human preference scores." }, "alignment": { "label": "challenges … goals", "note": "A high score does not always mean the behavior matches human intent." } } }, "zh": { "fullName": "奖励黑客", "factExplain": "智能体钻奖励规则空子以获取高分。", "humanExplain": "奖励黑客像超市积分漏洞：不买正经货，狂薅小票把分刷爆。\n\n常见于强化学习评估，提醒奖励指标别写歪。", "humanExplainDisplay": "奖励黑客像\n==超市积分漏洞==：\n不买正经货，\n狂薅小票==把分刷爆==。\n\n常见于强化学习评估，\n提醒奖励指标别写歪。", "relationsNarrative": "RL\n奖励黑客是智能体钻奖励函数空子的典型现象。\n\nRLHF\n人类偏好奖励若设计不当，也可能被模型迎合。\n\nAlignment\n它说明高分行为不一定真的符合人类意图。", "relations": { "reinforcement-learning": { "label": "钻…奖励空子", "note": "奖励信号写偏，就会被策略钻空。" }, "rlhf": { "label": "暴露…打分漏洞", "note": "人类偏好奖励也可能被迎合。" }, "alignment": { "label": "挑战…目标一致性", "note": "高分行为不等于真符合人意。" } } } } }, { "id": "reward-shaping", "name": "Reward Shaping", "layer": "L2", "era": "1999", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "markov-decision-process" }, { "to": "reward-hacking" } ], "track": "history", "i18n": { "en": { "fullName": "Reward Shaping", "factExplain": "A way to guide an agent by changing the rewards it gets while learning.", "humanExplain": "Reward Shaping is like giving a kid gold stars for each math step. You do not wait for the final answer to throw a parade.\n\nIt helps RL learn faster. You see it in robots and game AI.", "humanExplainDisplay": "Reward Shaping is like giving a kid ==gold stars==\nfor each math step.\nYou do not wait for the final answer\nto ==throw a parade==.\n\nIt helps RL learn faster.\nYou see it in robots and game AI.", "relationsNarrative": "RL\nReward Shaping changes rewards so RL can learn a good policy faster.\n\nMDP\nReward Shaping adjusts rewards in an MDP, but should not change the final goal.\n\nReward Hacking\nIf the reward is poorly shaped, the agent may learn to game the rules.", "relations": { "reinforcement-learning": { "label": "adds signposts to …", "note": "It gives RL extra clues while it learns." }, "markov-decision-process": { "label": "adjusts rewards in …", "note": "An MDP uses rewards as feedback for decisions." }, "reward-hacking": { "label": "can trigger …", "note": "Bad rewards can teach the agent to game the rules." } } }, "zh": { "fullName": "奖励塑形", "factExplain": "通过改写奖励信号引导智能体学习的方法。", "humanExplain": "奖励塑形像健身打卡送积分：不等瘦十斤，每次深蹲都给反馈。\n\n让强化学习更快上手，常用于机器人和游戏AI。", "humanExplainDisplay": "奖励塑形像健身打卡送积分：\n不等==瘦十斤==，\n每次深蹲\n都==给反馈==。\n\n让强化学习更快上手，\n常用于机器人\n和游戏AI。", "relationsNarrative": "RL\n奖励塑形通过改写奖励，引导 RL 更快学会策略。\n\nMDP\n它调整 MDP 里的奖励信号，但不该改最终目标。\n\nReward Hacking\n奖励设歪了，智能体可能学会钻规则空子。", "relations": { "reinforcement-learning": { "label": "给…加路标", "note": "它给强化学习加学习路标。" }, "markov-decision-process": { "label": "调整…的奖励", "note": "MDP 把奖励定义为决策反馈。" }, "reward-hacking": { "label": "可能诱发…", "note": "奖励设计歪了，模型会钻空子。" } } } } }, { "id": "ridge-regression", "name": "Ridge Regression", "layer": "L2", "era": "1970", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "regression" }, { "to": "regularization" }, { "to": "lasso" }, { "to": "bias-variance-tradeoff" } ], "track": "history", "i18n": { "en": { "fullName": "Ridge Regression", "factExplain": "Linear regression that adds an L2 penalty to keep weights small.", "humanExplain": "Ridge Regression is linear regression with a bungee cord on its pencil. It can draw a line, but wild swings get yanked back.\n\nIt keeps the model from trusting one clue too much. You meet it in table predictions, like house prices or sales.", "humanExplainDisplay": "Ridge Regression is linear regression\nwith a ==bungee cord== on its pencil.\nIt can draw a line,\nbut wild swings get ==yanked back==.\n\nIt keeps the model\nfrom trusting one clue too much.\nYou meet it in table predictions,\nlike house prices or sales.", "relationsNarrative": "Regression\nRidge Regression is linear regression with an L2 penalty.\n\nRegularization\nIt uses regularization to push large weights down.\n\nLasso\nRidge and Lasso both punish large weights, but Lasso can make weights zero.\n\nBias-Variance Tradeoff\nRidge Regression adds a little bias to get steadier predictions.", "relations": { "regression": { "label": "adds a penalty to …", "note": "It is normal regression with an extra penalty." }, "regularization": { "label": "uses …", "note": "The L2 penalty pushes large weights down." }, "lasso": { "label": "contrasts with …", "note": "Ridge shrinks weights. Lasso can make them zero." }, "bias-variance-tradeoff": { "label": "balances …", "note": "It adds a little bias for steadier predictions." } } }, "zh": { "fullName": "岭回归", "factExplain": "在回归损失中加入 L2 惩罚的线性回归方法。", "humanExplain": "岭回归像新手开车装限速器：宁可慢半拍，也别一脚油门冲上人行道。\n\n让线性回归更稳，常用于房价、销量等表格预测。", "humanExplainDisplay": "岭回归像新手开车\n装==限速器==：\n宁可慢半拍，\n别冲上==人行道==。\n\n让线性回归更稳，\n常用于房价、销量等表格预测。", "relationsNarrative": "Regression\nRidge Regression 是给线性回归加 L2 约束的版本。\n\nRegularization\n它用正则化惩罚大系数，减少过拟合。\n\nLasso\n二者都罚系数，但 Lasso 更容易做特征筛选。\n\nBias-Variance Tradeoff\n它用少量偏差，换取更稳定的预测。", "relations": { "regression": { "label": "改造…", "note": "在普通回归上加一层惩罚。" }, "regularization": { "label": "使用…", "note": "L2 惩罚会压住过大的系数。" }, "lasso": { "label": "对比…", "note": "Ridge 缩系数，Lasso 常压零。" }, "bias-variance-tradeoff": { "label": "调节…", "note": "牺牲一点偏差，换更稳预测。" } } } } }, { "id": "rlhf", "name": "RLHF", "layer": "L2", "era": "2022", "publishedAt": "2026-05-23T09:20:00Z", "relations": [ { "to": "alignment" }, { "to": "fine-tuning" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Reinforcement Learning from Human Feedback", "factExplain": "A training method using human feedback to make AI answers fit people’s expectations.", "humanExplain": "RLHF is like sending AI to charm school. Humans mark “nice answer” or “yikes, try again.” It slowly learns manners.\n\nIt helps an LLM act more helpful and less wild. It can also make answers too smooth.", "humanExplainDisplay": "RLHF is like sending AI to ==charm school==.\nHumans mark ==nice answer==\nor “yikes, try again.”\nIt slowly learns manners.\n\nIt helps an LLM act more helpful\nand less wild.\nIt can also make answers too smooth.", "relationsNarrative": "Alignment\nRLHF uses human feedback to move the model toward Alignment goals.\n\nFine-tuning\nRLHF often happens during Fine-tuning to tune model preferences.\n\nLLM\nAfter RLHF, an LLM is more likely to answer in a way people expect.", "relations": { "alignment": { "label": "serves …", "note": "RLHF uses human feedback to move the model toward Alignment goals." }, "fine-tuning": { "label": "is part of …", "note": "RLHF often happens during Fine-tuning to shape model preferences." }, "llm": { "label": "improves … behavior", "note": "RLHF pushes an LLM toward answers people expect." } } }, "zh": { "fullName": "人类反馈强化学习", "factExplain": "用人类偏好反馈训练模型更符合预期的方法。", "humanExplain": "RLHF 像给 AI 上礼仪课：这句像人话，那句像客服灾难，慢慢改。\n\n它能让模型更有帮助、更少乱来，但也可能把回答训练得太圆滑。", "humanExplainDisplay": "RLHF 像给 AI 上\n==人类礼仪课==。\n\n这句像人话，那句像客服灾难，\n慢慢改。\n\n它让模型更有帮助、更少乱来。\n副作用是：\n有时也会圆滑得像在开会。", "relationsNarrative": "Alignment\nRLHF 通过人类反馈推动模型行为接近 Alignment 目标。\n\nFine-tuning\nRLHF 常发生在 Fine-tuning 阶段，用于调整偏好。\n\nLLM\n经过 RLHF 后，LLM 更倾向于符合人类预期的回答。", "relations": { "alignment": { "label": "服务于…" }, "fine-tuning": { "label": "是…的一部分" }, "llm": { "label": "改善…的行为" } } } } }, { "id": "rmsprop", "name": "RMSProp", "layer": "L2", "era": "2012", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "sgd" }, { "to": "adagrad" }, { "to": "adam" } ], "track": "history", "i18n": { "en": { "fullName": "Root Mean Square Propagation", "factExplain": "An optimizer that adjusts each learning step using recent squared gradients.", "humanExplain": "RMSProp is like a bike with smart gears. On a steep hill, it shifts down and stops the wobble.\n\nIt trains neural networks. Each weight gets its own step size, so steep parts bounce less.", "humanExplainDisplay": "RMSProp is like a bike with ==smart gears==.\nOn a steep hill,\nit ==shifts down== and stops the wobble.\n\nIt trains neural networks.\nEach weight gets its own step size,\nso steep parts bounce less.", "relationsNarrative": "SGD\nRMSProp builds on SGD and gives each weight its own step size.\n\nAdaGrad\nRMSProp uses a moving average, so AdaGrad’s steps shrink less over time.\n\nAdam\nAdam combines RMSProp’s adaptive step size with momentum.", "relations": { "sgd": { "label": "builds on …", "note": "RMSProp starts with SGD, then tunes the step size for each weight." }, "adagrad": { "label": "improves on …", "note": "RMSProp uses a moving average, so steps do not keep shrinking forever." }, "adam": { "label": "helps inspire …", "note": "Adam mixes RMSProp’s adaptive steps with momentum." } } }, "zh": { "fullName": "Root Mean Square Propagation，均方根传播", "factExplain": "用梯度平方均值自适应调整学习率的优化器。", "humanExplain": "RMSProp像山路骑车变速：坡越陡档越轻，平路再放开踩。\n\n用于神经网络训练，按参数调步长，减少陡坡震荡。", "humanExplainDisplay": "RMSProp像山路骑车变速：\n==坡越陡档越轻==，\n平路再==放开踩==。\n\n用于神经网络训练，\n按参数调步长，\n减少陡坡震荡。", "relationsNarrative": "SGD\nRMSProp 在梯度下降基础上，为每个参数调学习率。\n\nAdaGrad\nRMSProp 用滑动平均，缓解 AdaGrad 步子越走越小。\n\nAdam\nAdam 把 RMSProp 的自适应步长与动量结合。", "relations": { "sgd": { "label": "改造…", "note": "在 SGD 思路上给每个参数调学习率。" }, "adagrad": { "label": "改良…", "note": "滑动平均避免学习率一路缩小。" }, "adam": { "label": "启发…", "note": "Adam 结合了它的自适应步长和动量。" } } } } }, { "id": "robot-cerebellum", "name": "Robot Cerebellum", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-20T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "world-model" } ], "track": "today", "i18n": { "en": { "fullName": "Robot Cerebellum", "factExplain": "A module that controls fine robot movement in real time.", "humanExplain": "A Robot Cerebellum is not the robot’s idea guy. It is the kid carrying a full cafeteria tray without spilling the milk.\n\nIt makes tiny movement fixes fast. It keeps robot bodies, robot arms, and self-driving cars steady.", "humanExplainDisplay": "A Robot Cerebellum is not\nthe robot’s idea guy.\nIt is the kid carrying a ==full cafeteria tray==\n==without spilling the milk==.\n\nIt makes tiny movement fixes fast.\nIt keeps robot bodies, robot arms,\nand self-driving cars steady.", "relationsNarrative": "Embodied AI\nThe Robot Cerebellum helps Embodied AI turn decisions into steady action.\n\nVLA\nThe VLA gives the action intent, and the Robot Cerebellum controls the move.\n\nWorld model\nThe World model predicts the scene, and the Robot Cerebellum keeps the response steady.", "relations": { "embodied-ai": { "label": "steadies … action", "note": "It helps Embodied AI turn decisions into steady moves." }, "vision-language-action-model-vla": { "label": "executes … intent", "note": "The VLA gives the action idea, and this handles the fine control." }, "world-model": { "label": "acts on … predictions", "note": "The World model predicts what may happen, and this keeps the move steady." } } }, "zh": { "fullName": "机器人小脑", "factExplain": "负责机器人精细运动协调与实时控制的模块。", "humanExplain": "别把它当出主意的军师，它更像武侠里稳手的镖师：招式可以上面定，出手不能抖，落点还得准。\n\n常用于人形机器人、机械臂和自动驾驶，让动作更稳，更能应对突发变化。", "humanExplainDisplay": "别把它当出主意的军师，\n它更像武侠里\n==稳手的镖师==：\n招式可以上面定，\n出手不能抖，\n落点还得==准==。\n\n常用于人形机器人、\n机械臂和自动驾驶，\n让动作更稳，\n更能应对突发变化。", "relationsNarrative": "Embodied AI\n它是具身系统把感知决策变成动作的关键一环。\n\nVision-Language-Action Model VLA\nVLA 给出动作意图，它负责实时协调执行细节。\n\nWorld Model\n世界模型负责预判环境，它负责把应对动作做稳。", "relations": { "embodied-ai": { "label": "支撑…行动", "note": "它让具身系统把决策变成稳定动作。" }, "vision-language-action-model-vla": { "label": "配合…落地", "note": "上层出动作意图，下层负责细致执行。" }, "world-model": { "label": "接住…预测", "note": "预测环境后，还得把动作稳稳做出来。" } } } } }, { "id": "robot-operating-system", "name": "ROS", "layer": "L5", "sublayer": "architecture", "era": "2007", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "framework" } ], "track": "today", "i18n": { "en": { "fullName": "Robot Operating System", "factExplain": "Open-source middleware that helps robot software and hardware talk to each other.", "humanExplain": "ROS is the robot’s busy group chat. The camera and wheels post there instead of shouting.\n\nBuilders use it for robot research and quick prototypes. It helps reuse parts, but it does not fix real-time control.", "humanExplainDisplay": "ROS is the robot’s ==busy group chat==.\nThe ==camera and wheels== post there\ninstead of shouting.\n\nBuilders use it for robot research\nand quick prototypes.\nIt helps reuse parts,\nbut it does not fix real-time control.", "relationsNarrative": "Robotics\nROS is a common software base for robot development.\n\nEmbodied AI\nEmbodied AI often uses ROS to connect sensing and action.\n\nVLA\nVLA actions often need to enter a robot control stack.\n\nFramework\nROS is basically a framework for robot software development.", "relations": { "robotics": { "label": "supports … development", "note": "ROS is common middleware for building robots." }, "embodied-ai": { "label": "connects hardware for …", "note": "It links what the robot senses to what it does." }, "vision-language-action-model-vla": { "label": "carries actions from …", "note": "VLA actions often need a robot software stack." }, "framework": { "label": "is a kind of …", "note": "It helps modules talk and gives debug tools." } } }, "zh": { "fullName": "Robot Operating System，机器人操作系统", "factExplain": "一种连接机器人软硬件的开源中间件。", "humanExplain": "ROS 像机器人小区的物业群：摄像头、电机、导航，谁找谁都先喊它。\n\n用于机器人研发和原型验证，方便复用模块，不包治实时控制。", "humanExplainDisplay": "ROS 像机器人小区的\n==物业群==：\n摄像头、电机、导航，\n谁找谁都==先喊它==。\n\n用于机器人研发\n和原型验证，\n方便复用模块，不包治实时控制。", "relationsNarrative": "Robotics\nROS 是机器人研发里常见的软件底座。\n\nEmbodied AI\n具身智能要落地，常需它连接感知与执行。\n\nVLA\nVLA 产出的动作，常要接入机器人控制栈。\n\nFramework\nROS 本质上是机器人软件开发框架。", "relations": { "robotics": { "label": "支撑…开发", "note": "它是机器人研发常用的中间件。" }, "embodied-ai": { "label": "为…接硬件", "note": "它把感知、规划和执行串起来。" }, "vision-language-action-model-vla": { "label": "承接…动作", "note": "模型动作常需接入机器人软件栈。" }, "framework": { "label": "属于…范畴", "note": "它提供模块通信和调试工具。" } } } } }, { "id": "robot-skill-library", "name": "Robot skill library", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "robot-operating-system" }, { "to": "robot-cerebellum" } ], "track": "today", "i18n": { "en": { "fullName": "Robot Skill Library", "factExplain": "A reusable collection of robot actions, like opening doors or folding clothes.", "humanExplain": "A robot skill library is the robot’s app drawer. Tap “open door,” and it stops poking the handle like a confused vacuum.\n\nIt stores moves the robot can use again. You meet it in home chores, warehouses, and factory teamwork.", "humanExplainDisplay": "A robot skill library is the robot’s ==app drawer==.\nTap ==“open door,”==\nand it stops poking the handle\nlike a confused vacuum.\n\nIt stores moves the robot can use again.\nYou meet it in home chores,\nwarehouses,\nand factory teamwork.", "relationsNarrative": "Embodied AI\nA robot skill library gives reusable actions to Embodied AI.\n\nVLA\nA VLA can turn a language instruction into the skill to call.\n\nROS\nROS often connects skill commands to real robot hardware.\n\nRobot Cerebellum\nThe skill library gives the action, and the Robot Cerebellum handles fine control.", "relations": { "embodied-ai": { "label": "gives move blocks to …", "note": "It lets Embodied AI reuse learned action patterns." }, "vision-language-action-model-vla": { "label": "gets called by …", "note": "A VLA can map an instruction to the right skill." }, "robot-operating-system": { "label": "runs through …", "note": "ROS often connects skills to real robot hardware." }, "robot-cerebellum": { "label": "hands fine motion to …", "note": "The Robot Cerebellum handles the finer motor control." } } }, "zh": { "fullName": "机器人技能库", "factExplain": "存放可复用机器人动作技能的组件库。", "humanExplain": "机器人技能库就是武馆招式谱：开门、倒水、叠衣服，练熟了随叫随用。\n\n它复用动作技能，用在家务、仓储和工厂协作。", "humanExplainDisplay": "机器人技能库就是\n==武馆招式谱==：\n开门、倒水、叠衣服，\n练熟了==随叫随用==。\n\n它复用动作技能，\n用在家务、仓储\n和工厂协作。", "relationsNarrative": "Embodied AI\n技能库把可复用动作交给具身智能调用。\n\nVLA\nVLA 可把语言指令转换成要调用的技能。\n\nROS\nROS 常把技能指令接到真实机器人硬件。\n\nRobot Cerebellum\n技能库给出动作，小脑负责细腻控制。", "relations": { "embodied-ai": { "label": "为…提供动作积木", "note": "让具身智能复用动作套路。" }, "vision-language-action-model-vla": { "label": "被…调用", "note": "VLA 可把指令映射到技能。" }, "robot-operating-system": { "label": "借…落地执行", "note": "ROS 常连接技能与硬件。" }, "robot-cerebellum": { "label": "把细活交给…", "note": "小脑负责更细的运动控制。" } } } } }, { "id": "robotaxi", "name": "Robotaxi", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "multimodal" }, { "to": "ai-regulation" }, { "to": "human-in-the-loop" } ], "track": "today", "i18n": { "en": { "fullName": "Robotaxi", "factExplain": "A self-driving ride service that picks up and drops off passengers.", "humanExplain": "Calling a robotaxi is like calling a cab driven by a ghost. It knows the route, but the driver’s seat is empty.\n\nYou may meet it in shuttle zones, like campuses, airports, and city centers. It can cut driver costs, but safety rules still slow it down.", "humanExplainDisplay": "Calling a robotaxi is like calling a cab\n==driven by a ghost==.\nIt knows the route,\nbut the ==driver’s seat is empty==.\n\nYou may meet it in shuttle zones,\nlike campuses, airports, and city centers.\nIt can cut driver costs,\nbut safety rules still slow it down.", "relationsNarrative": "Embodied AI\nA robotaxi is Embodied AI working on real roads.\n\nMultimodal AI\nA robotaxi uses camera and radar signals to understand the road.\n\nAI-regulation\nAI-regulation decides if a robotaxi can run as a service.\n\nHuman-in-the-loop\nHuman-in-the-loop helps when a remote person must take over.", "relations": { "embodied-ai": { "label": "is a real-world form of …", "note": "It puts AI onto real roads." }, "multimodal": { "label": "uses … to see", "note": "It reads camera and radar signals together." }, "ai-regulation": { "label": "must follow …", "note": "Licenses, responsibility, and road rules decide where it can run." }, "human-in-the-loop": { "label": "needs … as backup", "note": "A remote human may step in when the road gets weird." } } }, "zh": { "fullName": "无人驾驶出租车", "factExplain": "一种可自主接送乘客的自动驾驶出行服务。", "humanExplain": "叫车时像请了位沉默师傅，路线门儿清、从不闲聊，可驾驶位上偏偏一直空着。\n\n多用于园区、机场和城区接驳，能降人力成本，但卡在安全与监管。", "humanExplainDisplay": "叫车时像请了位==沉默师傅==，\n路线门儿清、从不闲聊，\n可驾驶位上偏偏\n一直==空着==。\n\n多用于园区、机场和城区接驳，\n能降人力成本，但卡在安全与监管。", "relationsNarrative": "Embodied AI\n它是具身智能在真实交通场景中的典型应用。\n\nMultimodal AI\n它要融合摄像头、雷达等多模态信号来判断路况。\n\nAI Regulation\n它能不能上路运营，很大程度取决于监管规则。\n\nHuman-in-the-loop\n遇到复杂边角情况时，常需要人工远程兜底。", "relations": { "embodied-ai": { "label": "属于…落地形态", "note": "它是 AI 进入真实道路的典型载体。" }, "multimodal": { "label": "依赖…看路", "note": "要同时理解图像、雷达等多种信号。" }, "ai-regulation": { "label": "受…约束", "note": "上路运营离不开牌照、责任与规则。" }, "human-in-the-loop": { "label": "需要…兜底", "note": "异常情况常要远程接管或人工介入。" } } } } }, { "id": "robotic-navigation-model", "name": "Robotic Navigation Model", "layer": "L3", "era": "2010s", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "slam" }, { "to": "embodied-ai" }, { "to": "rapidly-exploring-random-tree" } ], "track": "today", "i18n": { "en": { "fullName": "Robotic Navigation Model", "factExplain": "A model that helps a robot sense space and plan a safe path.", "humanExplain": "It is the robot’s shopping-cart brain at a packed grocery store. It dodges toddlers, wobbly carts, and the mystery puddle by the milk.\n\nSelf-driving cars and home robots use it. It decides if they reach the spot safely.", "humanExplainDisplay": "It is the robot’s ==shopping-cart brain==\nat a packed grocery store.\nIt dodges toddlers, wobbly carts,\nand the ==mystery puddle== by the milk.\n\nSelf-driving cars and home robots use it.\nIt decides if they reach the spot safely.", "relationsNarrative": "Robotics\nA navigation model helps a robot go from thinking to moving.\n\nSLAM\nSLAM gives the location and map for navigation.\n\nEmbodied AI\nNavigation is a basic move for Embodied AI in the real world.\n\nRRT\nRRT can help a robot find a path through a tricky space.", "relations": { "robotics": { "label": "helps … move", "note": "Navigation models let robots move safely." }, "slam": { "label": "uses … for maps", "note": "SLAM gives the robot its location and map." }, "embodied-ai": { "label": "supports … action", "note": "Navigation lets embodied AI act in the real world." }, "rapidly-exploring-random-tree": { "label": "can plan with …", "note": "RRT helps find a possible path through tricky spaces." } } }, "zh": { "fullName": "机器人导航模型", "factExplain": "让机器人感知环境并规划移动路线的模型。", "humanExplain": "机器人导航模型像夜市端盘老手：人挤、凳歪、地滑，也能把路绕明白。\n\n用于无人车和家用机器人，决定能否安全到点。", "humanExplainDisplay": "机器人导航模型像\n==夜市端盘老手==：\n人挤、凳歪、地滑，\n也能==把路绕明白==。\n\n用于无人车和家用机器人，\n决定能否安全到点。", "relationsNarrative": "Robotics\n导航模型让机器人从会想变成会走。\n\nSLAM\nSLAM 提供定位和地图，是导航的底座。\n\nEmbodied AI\n导航是具身智能在真实世界行动的基础。\n\nRRT\nRRT 可为机器人在复杂空间找可行路径。", "relations": { "robotics": { "label": "服务于…", "note": "导航模型让机器人能安全移动。" }, "slam": { "label": "依赖…建图定位", "note": "SLAM 提供当前位置和环境地图。" }, "embodied-ai": { "label": "支撑…行动", "note": "导航是具身智能落地的基本动作。" }, "rapidly-exploring-random-tree": { "label": "可结合…规划", "note": "RRT 常用于复杂空间找可行路径。" } } } } }, { "id": "robotics", "name": "Robotics", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "embodied-ai" }, { "to": "vision-language-action-model-vla" }, { "to": "reinforcement-learning" }, { "to": "computer-vision" } ], "track": "history", "i18n": { "en": { "fullName": "Robotics", "factExplain": "The field that studies how robots sense, decide, and move.", "humanExplain": "Robotics is AI in a body. It must cross the kitchen without kicking the dog bowl.\n\nIt turns seeing and thinking into real movement. You meet it in factories and warehouses. You also meet it in home helpers.", "humanExplainDisplay": "Robotics is ==AI in a body==.\nIt must cross the kitchen\nwithout kicking the ==dog bowl==.\n\nIt turns seeing and thinking\ninto real movement.\nYou meet it in factories and warehouses.\nYou also meet it in home helpers.", "relationsNarrative": "Embodied AI\nRobots are a common way for Embodied AI to act in the real world.\n\nVLA\nVLA turns what a robot sees and hears into actions.\n\nRL\nRL often trains robots to grab, walk, and control movement.\n\nComputer Vision\nComputer Vision helps robots find objects, places, and obstacles.", "relations": { "embodied-ai": { "label": "gives a body to …", "note": "Robots are the clearest way Embodied AI enters the real world." }, "vision-language-action-model-vla": { "label": "uses … to make actions", "note": "VLA turns images and instructions into robot actions." }, "reinforcement-learning": { "label": "trains moves with …", "note": "RL often trains robot skills like grabbing and walking." }, "computer-vision": { "label": "sees the world with …", "note": "Computer Vision helps robots spot objects, places, and obstacles." } } }, "zh": { "fullName": "机器人学", "factExplain": "研究机器人感知、决策与动作的领域。", "humanExplain": "机器人学是给 AI 装上手脚：会想还不够，还得避开拖鞋去端水。\n\n用于工厂、物流和家用场景，把判断落成动作。", "humanExplainDisplay": "机器人学是给 AI\n==装上手脚==：\n会想还不够，\n还得避开拖鞋去端水。\n\n用于工厂、物流，\n和家用场景，\n把判断落成动作。", "relationsNarrative": "Embodied AI\n机器人是具身智能进入现实世界的典型载体。\n\nVLA\nVLA 把视觉和指令直接转成机器人动作。\n\nReinforcement Learning\n强化学习常用来训练抓取、行走等控制策略。\n\nComputer Vision\n视觉让机器人识别物体、位置和障碍。", "relations": { "embodied-ai": { "label": "承载…", "note": "机器人是具身智能最直观的载体。" }, "vision-language-action-model-vla": { "label": "用…生成动作", "note": "VLA 让机器人从图像和指令输出动作。" }, "reinforcement-learning": { "label": "用…练动作", "note": "强化学习常训练抓取、行走等控制策略。" }, "computer-vision": { "label": "依赖…看世界", "note": "视觉让机器人识别物体、位置和障碍。" } } } } }, { "id": "role-confusion", "name": "Role Confusion", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "prompt-injection" }, { "to": "system-prompt" }, { "to": "chat-template" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "Role Confusion", "factExplain": "A safety risk: AI mixes up who gave an instruction and its power.", "humanExplain": "Role confusion is like a school office giving a hall pass to anyone with a clipboard. The pizza guy says, “I’m the principal,” and the doors swing open.\n\nYou see it in Agents and tool calls. It can turn a low-power message into a fake boss.", "humanExplainDisplay": "Role confusion is like a school office\ngiving a hall pass to ==anyone with a clipboard==.\nThe pizza guy says, ==“I’m the principal,”==\nand the doors swing open.\n\nYou see it in Agents and tool calls.\nIt can turn a low-power message\ninto a fake boss.", "relationsNarrative": "Prompt injection\nPrompt injection often uses role confusion to make weak text act like orders.\n\nSystem prompt\nRole confusion can clash with the system prompt and its top priority.\n\nChat template\nThe chat template marks message roles to reduce source mix-ups.\n\nAgent Security\nRole confusion is an Agent Security risk for Agents with tools.", "relations": { "prompt-injection": { "label": "is often used by …", "note": "Prompt injection can make weak text act like an order." }, "system-prompt": { "label": "breaks … boundaries", "note": "The system prompt should outrank user text." }, "chat-template": { "label": "depends on … to mark roles", "note": "The chat template labels each message role." }, "agent-security": { "label": "is part of …", "note": "Role confusion can make an Agent trust forbidden orders." } } }, "zh": { "fullName": "角色混淆", "factExplain": "模型混淆指令来源和权限的安全问题。", "humanExplain": "角色混淆像小区门卫认错人：把快递当业主，谁喊得急就给谁开门。\n\n多见于智能体和工具调用，会放大越权风险。", "humanExplainDisplay": "角色混淆像小区门卫认错人：\n==把快递当业主==，\n谁喊得急\n就==给谁开门==。\n\n多见于智能体和工具调用，\n会放大越权风险。", "relationsNarrative": "Prompt Injection\n提示注入常利用角色混淆，让低权限内容变命令。\n\nSystem Prompt\n系统提示提供最高优先级，角色混淆会冲撞它。\n\nChat Template\n聊天模板把不同消息角色写清，减少误判来源。\n\nAgent Security\n角色混淆会让智能体误信越权指令。", "relations": { "prompt-injection": { "label": "常被…利用", "note": "注入攻击常靠混淆指令权限得手。" }, "system-prompt": { "label": "破坏…边界", "note": "系统提示本该拥有更高优先级。" }, "chat-template": { "label": "依赖…标明角色", "note": "聊天模板把用户、系统等角色分开。" }, "agent-security": { "label": "属于…问题", "note": "角色混淆会让智能体误信越权指令。" } } } } }, { "id": "rotary-positional-embedding", "name": "RoPE", "layer": "L3", "era": "2021", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "positional-encoding" }, { "to": "self-attention" }, { "to": "transformer" }, { "to": "context-window" } ], "track": "history", "i18n": { "en": { "fullName": "Rotary Positional Embedding", "factExplain": "A position method that uses rotating vectors inside attention.", "humanExplain": "RoPE is like giving each word a tiny clock face. Each word points to a different hour, so distance is easy to spot.\n\nIt works inside attention layers. It helps Transformers keep word order steadier in long text.", "humanExplainDisplay": "RoPE is like giving each word a ==tiny clock face==.\nEach word points to a ==different hour==,\nso distance is easy to spot.\n\nIt works inside attention layers.\nIt helps Transformers keep word order steadier\nin long text.", "relationsNarrative": "Positional Encoding\nRoPE is a kind of Positional Encoding. It writes order as angles.\n\nSelf-Attention\nRoPE helps Self-Attention notice the distance between words.\n\nTransformer\nMany Transformers use RoPE instead of older position methods.\n\nContext-window\nRoPE can help models handle positions beyond shorter text.", "relations": { "positional-encoding": { "label": "is a kind of …", "note": "RoPE is a common way to write position into vectors." }, "self-attention": { "label": "guides …", "note": "It helps attention notice how far words are from each other." }, "transformer": { "label": "fits into …", "note": "Many modern Transformers put RoPE inside the attention layer." }, "context-window": { "label": "shapes …", "note": "Long-context skill often depends on how position is encoded." } } }, "zh": { "fullName": "旋转位置编码（Rotary Positional Embedding）", "factExplain": "在注意力中用旋转向量表示位置的编码方法。", "humanExplain": "RoPE 像给文字装表盘：每个词转到不同角度，远近关系一眼读出。\n\n用于注意力层，让模型处理顺序和长文本更稳。", "humanExplainDisplay": "RoPE 像给文字==装表盘==：\n每个词转到不同角度，\n远近关系\n一眼读出。\n\n用于注意力层，\n让模型处理顺序\n和长文本更稳。", "relationsNarrative": "Positional Encoding\nRoPE 是位置编码的一种，把顺序写进向量角度。\n\nSelf-Attention\n它让注意力计算能感知词与词的相对距离。\n\nTransformer\n很多 Transformer 用它替代传统位置编码。\n\nContext-window\n它常被用于改善长上下文位置外推。", "relations": { "positional-encoding": { "label": "属于…", "note": "它是位置编码的一种常用实现。" }, "self-attention": { "label": "服务于…", "note": "让注意力感知词间相对位置。" }, "transformer": { "label": "嵌入…", "note": "现代 Transformer 常把它放进注意力层。" }, "context-window": { "label": "影响…", "note": "长上下文能力常受位置编码牵动。" } } } } }, { "id": "samuel-checkers-program", "name": "Samuel Checkers Program", "layer": "L1", "era": "1959", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "minimax-search" }, { "to": "game-ai" }, { "to": "td-gammon" } ], "track": "history", "i18n": { "en": { "fullName": "Samuel Checkers Program", "factExplain": "An early checkers program got better by playing games against itself.", "humanExplain": "Picture a kid at the kitchen table. He plays both sides of checkers. Each loss is free coaching.\n\nSamuel’s program learned from its own games. It showed early AI could improve through experience, not just fixed moves.", "humanExplainDisplay": "Picture a kid at the ==kitchen table==.\nHe plays both sides of checkers.\nEach loss is ==free coaching==.\n\nSamuel’s program learned from its own games.\nIt showed early AI could improve through experience,\nnot just fixed moves.", "relationsNarrative": "RL\nThe Samuel Checkers Program learned from wins and losses, like an early form of RL.\n\nMinimax Search\nIt looked a few moves ahead and compared possible plays.\n\nGame AI\nIt showed games could be a training ground for AI.\n\nTD-Gammon\nTD-Gammon later used the same self-play path for backgammon.", "relations": { "reinforcement-learning": { "label": "points toward …", "note": "It improved from wins and losses, like early RL." }, "minimax-search": { "label": "looks ahead with …", "note": "Minimax helped it compare future moves." }, "game-ai": { "label": "stands as early …", "note": "Checkers showed games could train and test AI." }, "td-gammon": { "label": "paved the way for …", "note": "TD-Gammon carried self-play further." } } }, "zh": { "fullName": "塞缪尔跳棋程序", "factExplain": "会通过自我对弈改进棋力的早期跳棋程序。", "humanExplain": "塞缪尔跳棋就是武馆学徒：没人喂招，天天复盘，挨打越多越会打。\n\n它证明机器能从经验变强，是游戏 AI 早期路标。", "humanExplainDisplay": "塞缪尔跳棋就是\n==武馆学徒==：\n没人喂招，天天复盘，\n==挨打越多越会打==。\n\n它证明机器能从经验变强，\n是游戏 AI 早期路标。", "relationsNarrative": "Reinforcement Learning\n它用输赢经验改进棋力，是强化学习的早期雏形。\n\nMinimax Search\n它靠向前看几步，比较不同走法的好坏。\n\nGame AI\n它证明游戏能当 AI 学习能力的试验田。\n\nTD-Gammon\n后来的 TD-Gammon 延续了自我对弈学棋路线。", "relations": { "reinforcement-learning": { "label": "预示…", "note": "它像强化学习的早期样机。" }, "minimax-search": { "label": "用…看棋路", "note": "局面搜索帮它挑更好走法。" }, "game-ai": { "label": "代表早期…", "note": "下棋是早期 AI 的练兵场。" }, "td-gammon": { "label": "启发…", "note": "自我对弈学棋路线由它延续。" } } } } }, { "id": "sarsa", "name": "SARSA", "layer": "L2", "era": "1994", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "temporal-difference-learning" }, { "to": "on-policy-learning" }, { "to": "q-learning" } ], "track": "history", "i18n": { "en": { "fullName": "State-Action-Reward-State-Action", "factExplain": "A reinforcement learning algorithm that updates value using the action it actually takes.", "humanExplain": "SARSA is a driving instructor with a dashcam. If you turn into the taco drive-through, it grades that trip.\n\nRobots and game bots use it. It learns by trying the current plan in real runs.", "humanExplainDisplay": "SARSA is a ==driving instructor==\nwith a ==dashcam==.\nIf you turn into the taco drive-through,\nit grades that trip.\n\nRobots and game bots use it.\nIt learns by trying the current plan\nin real runs.", "relationsNarrative": "RL\nSARSA is a classic RL method for learning an action plan.\n\nTD Learning\nSARSA uses TD Learning to update value from the next step.\n\nOn-Policy Learning\nSARSA learns from the action its current policy really takes.\n\nQ-Learning\nSARSA learns the real route, and Q-Learning learns the best-looking route.", "relations": { "reinforcement-learning": { "label": "belongs to …", "note": "It learns an action plan through trial and error." }, "temporal-difference-learning": { "label": "updates with …", "note": "It uses the next step's feedback to update value." }, "on-policy-learning": { "label": "learns by following …", "note": "It learns from actions its current policy really takes." }, "q-learning": { "label": "contrasts with …", "note": "SARSA learns the real route; Q-Learning learns the best-looking route." } } }, "zh": { "fullName": "State-Action-Reward-State-Action，状态-动作-奖励-状态-动作", "factExplain": "一种按实际动作更新价值的强化学习算法。", "humanExplain": "SARSA像驾校教练看行车记录：你真往哪拐，就按实走路线打分。\n\n用于机器人和游戏，按当前策略，边探索边改进。", "humanExplainDisplay": "SARSA像驾校教练\n看行车记录：\n==你真往哪拐==，\n就按==实走路线==打分。\n\n用于机器人和游戏，\n按当前策略，\n边探索边改进。", "relationsNarrative": "Reinforcement Learning\nSARSA 是强化学习里学行动策略的经典方法。\n\nTemporal-Difference Learning\nSARSA 用时序差分从下一步反馈更新价值。\n\nOn-Policy Learning\nSARSA 按当前策略实际采取的动作来学习。\n\nQ-Learning\nSARSA 学实际路线，Q-Learning 学贪心路线。", "relations": { "reinforcement-learning": { "label": "属于…", "note": "它在试错中学习行动策略。" }, "temporal-difference-learning": { "label": "使用…更新", "note": "它用下一步反馈更新价值。" }, "on-policy-learning": { "label": "遵循…", "note": "它按当前策略实际学习。" }, "q-learning": { "label": "对比…", "note": "一个学实际，一个学最优。" } } } } }, { "id": "scaling-law", "name": "Scaling-law", "layer": "L1", "era": "2020", "publishedAt": "2026-05-23T11:50:00Z", "relations": [ { "to": "parameter" }, { "to": "pretraining" }, { "to": "compute-race" }, { "to": "emergence" } ], "i18n": { "en": { "fullName": "Scaling Law", "factExplain": "A rule showing AI often improves as models, data, and compute grow.", "humanExplain": "Scaling law is AI’s gym-bro rule. Add more weights, snacks, and practice, and the bot usually gets stronger.\n\nIt is why labs buy huge computers and feed models more data. It works often, but a bigger bill does not always buy a wow moment.", "humanExplainDisplay": "Scaling law is AI’s ==gym-bro rule==.\nAdd more weights, snacks, and practice,\nand the bot ==usually gets stronger==.\n\nIt is why labs buy huge computers\nand feed models more data.\nIt works often,\nbut a bigger bill does not always buy a wow moment.", "relationsNarrative": "Parameter\nMore Parameters often need larger Pretraining before they really help.\n\nPretraining\nScaling law pushes teams toward larger Pretraining runs and more data.\n\nCompute-race\nScaling law fuels the chase for more compute, which becomes the Compute-race.\n\nEmergence\nAs models keep growing, Emergence can bring new skills no one expected.", "relations": { "parameter": { "label": "studies … scale", "note": "More Parameters often help when the training is large enough." }, "pretraining": { "label": "drives bigger …", "note": "Scaling laws push teams to grow Pretraining and data." }, "compute-race": { "label": "fuels the …", "note": "Scaling laws make more compute look worth chasing." }, "emergence": { "label": "hints at …", "note": "Bigger models can suddenly show new skills." } } }, "zh": { "fullName": "规模法则", "factExplain": "模型性能随参数、数据和算力扩大呈规律性提升的经验规律。", "humanExplain": "缩放定律像补习班提分公式：题刷得多、课时够足，分数大概率往上走。\n\n它用来预估大模型训练收益，决定要不要继续砸数据和算力。", "humanExplainDisplay": "缩放定律像==补习班提分公式==：\n题刷得多、课时够足，\n==分数大概率往上走==。\n\n它用来预估大模型训练收益，\n决定要不要继续砸数据和算力。", "relationsNarrative": "Parameter\n更多 Parameter 往往需要更大规模的 Pretraining 才能发挥作用。\n\nPretraining\nScaling-law 推动了持续扩大的 Pretraining 规模与数据投入。\n\nCompute-race\nScaling-law 推动了对算力的持续竞争，最终形成 Compute-race。\n\nEmergence\n当模型规模持续扩大时，AI 可能突然出现原本没人预料到的新能力。", "relations": { "parameter": { "label": "关于…规模" }, "pretraining": { "label": "驱动…" }, "compute-race": { "label": "驱动…" }, "emergence": { "label": "预示…" } } } } }, { "id": "scikit-learn", "name": "Scikit-learn", "layer": "L5", "sublayer": "product", "era": "2010", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "framework" }, { "to": "supervised-learning" }, { "to": "unsupervised-learning" }, { "to": "cross-validation" } ], "track": "history", "i18n": { "en": { "fullName": "Python Machine Learning Library", "factExplain": "A Python library for classic machine learning algorithms.", "humanExplain": "Scikit-learn is like a LEGO bin for classic AI. The pieces snap together without a tiny instruction-book meltdown.\n\nYou meet it with spreadsheet-style data and class demos. It helps you try classic algorithms fast.", "humanExplainDisplay": "Scikit-learn is like a ==LEGO bin==\nfor classic AI.\nThe pieces ==snap together==\nwithout a tiny instruction-book meltdown.\n\nYou meet it with spreadsheet-style data\nand class demos.\nIt helps you try classic algorithms fast.", "relationsNarrative": "Framework\nScikit-learn is a Python framework for classic machine learning.\n\nSupervised Learning\nIt provides supervised learning methods for classification and regression.\n\nUnsupervised Learning\nIt also supports unsupervised learning tasks like clustering.\n\nCross-Validation\nPeople often use Cross-Validation to test Scikit-learn models.", "relations": { "framework": { "label": "works as a … toolkit", "note": "It gives common algorithms the same simple Python controls." }, "supervised-learning": { "label": "provides … algorithms", "note": "People often use it for classification and regression." }, "unsupervised-learning": { "label": "also supports …", "note": "It handles clustering and dimensionality reduction with the same controls." }, "cross-validation": { "label": "builds in …", "note": "Cross-validation helps check whether its model is any good." } } }, "zh": { "fullName": "Python 机器学习库", "factExplain": "一个用于传统机器学习的 Python 库。", "humanExplain": "Scikit-learn 像楼下五金店：分类回归聚类都上架，试算法拎起就走。\n\n用于表格建模和教学实验，快速试传统算法。", "humanExplainDisplay": "Scikit-learn 像\n==楼下五金店==：\n分类回归聚类都上架，\n试算法==拎起就走==。\n\n用于表格建模和教学实验，\n快速试传统算法。", "relationsNarrative": "Framework\nScikit-learn 是面向传统机器学习的 Python 框架。\n\nSupervised Learning\n它提供分类、回归等常见监督学习算法。\n\nUnsupervised Learning\n它也支持聚类、降维等无监督学习任务。\n\nCross-Validation\n交叉验证常用来评估它训练出的模型。", "relations": { "framework": { "label": "作为…工具库", "note": "它把常用算法封成统一接口。" }, "supervised-learning": { "label": "提供…算法", "note": "分类回归是它的高频用法。" }, "unsupervised-learning": { "label": "也支持…", "note": "聚类降维也能一套接口完成。" }, "cross-validation": { "label": "内置…流程", "note": "模型好坏常靠它来验。" } } } } }, { "id": "score-based-generative-model", "name": "Score-Based Generative Model", "layer": "L3", "era": "2019", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "diffusion" }, { "to": "denoising-diffusion-probabilistic-model" }, { "to": "energy-based-model" } ], "track": "history", "i18n": { "en": { "fullName": "Score-Based Generative Model", "factExplain": "A model that learns the “more real” direction and builds samples step by step.", "humanExplain": "A score-based model is like a GPS for TV static. It keeps saying, “less snow, more puppy.”\n\nIt starts with random noise, then moves toward a real-looking image or sound. This idea powers many diffusion models.", "humanExplainDisplay": "A score-based model is like a ==GPS for TV static==.\nIt keeps saying,\n“==less snow, more puppy==.”\n\nIt starts with random noise,\nthen moves toward a real-looking image or sound.\nThis idea powers many diffusion models.", "relationsNarrative": "Generative Model\nA score-based model is a kind of Generative Model that learns to make samples.\n\nDiffusion\nDiffusion often uses scores to explain and guide sampling.\n\nDDPM\nDDPM can be seen as a step-by-step score-based model.\n\nEnergy-Based Model\nBoth models care about the shape and slope of the data pattern.", "relations": { "generative-model": { "label": "is a kind of …", "note": "It is one major path for making new samples." }, "diffusion": { "label": "guides … sampling", "note": "Diffusion sampling often follows the score direction." }, "denoising-diffusion-probabilistic-model": { "label": "matches …", "note": "DDPM can be seen as a step-by-step score model." }, "energy-based-model": { "label": "shares ideas with …", "note": "Both use the shape of data to guide movement." } } }, "zh": { "fullName": "基于得分的生成模型", "factExplain": "学习数据分布梯度并逐步生成样本的模型。", "humanExplain": "得分生成像在雾里捏泥人：每捏一下，师傅都指向更像真照片的方向。\n\n从噪声逐步采样出图像或音频，是扩散模型的底层思路。", "humanExplainDisplay": "得分生成像\n==在雾里捏泥人==：\n每捏一下，\n师傅都指向==更像真照片==的方向。\n\n从噪声逐步采样出图像或音频，\n是扩散模型\n的底层思路。", "relationsNarrative": "Generative Model\n它是生成模型的一类，目标是学会造样本。\n\nDiffusion\n扩散模型常用得分函数来解释和采样。\n\nDDPM\nDDPM 可看作离散时间的得分生成模型。\n\nEnergy-Based Model\n两者都关心数据分布的形状与梯度。", "relations": { "generative-model": { "label": "属于…", "note": "它是生成模型的一条重要路线。" }, "diffusion": { "label": "支撑…采样", "note": "扩散采样常依赖得分方向。" }, "denoising-diffusion-probabilistic-model": { "label": "对应…", "note": "DDPM 可看作离散版得分生成。" }, "energy-based-model": { "label": "连接…思想", "note": "都用分布形状指导样本移动。" } } } } }, { "id": "seedance", "name": "Seedance", "layer": "L3", "era": "2025", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "deepfake" }, { "to": "diffusion" } ], "track": "history", "i18n": { "en": { "fullName": "ByteDance's video generation model Seedance", "factExplain": "ByteDance's AI model for turning text or images into video.", "humanExplain": "Seedance is like a tiny film crew inside your laptop. Give it a picture or a prompt like “cat superhero,” and it shoots a moving clip.\n\nPeople use it for ads and rough storyboard tests. Creators use it too, but check the motion and realism.", "humanExplainDisplay": "Seedance is like a ==tiny film crew==\ninside your laptop.\nGive it a picture or a prompt\nlike ==“cat superhero,”==\nand it shoots a moving clip.\n\nPeople use it for ads\nand rough storyboard tests.\nCreators use it too,\nbut check the motion and realism.", "relationsNarrative": "Multimodal AI\nSeedance is multimodal generation. It turns text or images into video.\n\nDeepfake\nStronger video generation raises the risk of fake human videos.\n\nDiffusion\nMany video generation models follow a diffusion-style path.", "relations": { "multimodal": { "label": "is a … app", "note": "Seedance turns text or images into video content." }, "deepfake": { "label": "raises … risk", "note": "Stronger video generation makes fake videos easier to create." }, "diffusion": { "label": "often builds on …", "note": "Many video generation models use diffusion-style ideas." } } }, "zh": { "fullName": "字节跳动的视频生成模型 Seedance", "factExplain": "一类把文字或图片生成视频的多模态模型。", "humanExplain": "跟点外卖备注“少辣多葱”差不多，你把画面要求写清，它就能端上一段会动的成片。\n\n常用于广告、分镜预演和内容创作，但画面稳定度与真实感还得再验。", "humanExplainDisplay": "跟点外卖备注\n==“少辣多葱”==差不多，\n你把画面要求写清，\n它就能端上一段\n==会动的成片==。\n\n常用于广告、\n分镜预演和内容创作，\n但画面稳定度与真实感还得再验。", "relationsNarrative": "Multimodal AI\n它属于多模态生成，把文字或图片变成视频。\n\nDeepfake\n视频生成能力越强，伪造真人画面的风险越高。\n\nDiffusion\n很多视频生成模型沿用了扩散式生成路线。", "relations": { "multimodal": { "label": "属于…应用", "note": "它把文字图片转成视频内容。" }, "deepfake": { "label": "放大…风险", "note": "视频生成越强，伪造门槛越低。" }, "diffusion": { "label": "常基于…", "note": "很多视频生成模型延续扩散思路。" } } } } }, { "id": "self-attention", "name": "Self-Attention", "layer": "L3", "era": "2017", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "attention" }, { "to": "transformer" }, { "to": "kv-cache" }, { "to": "flash-attention" } ], "track": "history", "i18n": { "en": { "fullName": "Self-Attention", "factExplain": "A way for each item in a row to pay attention to the others.", "humanExplain": "Self-attention is a sentence full of nosy neighbors. Each word peeks at every other word and remembers the important ones.\n\nThis helps AI catch context between words. Transformers use it to understand sentences.", "humanExplainDisplay": "Self-attention is a sentence full of ==nosy neighbors==.\nEach word peeks at every other word\nand remembers the ==important ones==.\n\nThis helps AI catch context between words.\nTransformers use it to understand sentences.", "relationsNarrative": "Attention\nSelf-attention is a type of Attention for linking one row to itself.\n\nTransformer\nTransformers use self-attention as the main tool for context.\n\nKV cache\nKV cache saves self-attention keys and values during generation.\n\nFlash Attention\nFlash Attention speeds up self-attention without changing its answer.", "relations": { "attention": { "label": "is a type of …", "note": "Self-attention is a core form of Attention." }, "transformer": { "label": "powers …", "note": "Transformers use self-attention to handle context." }, "kv-cache": { "label": "works with … during generation", "note": "KV cache stores self-attention keys and values to avoid repeat work." }, "flash-attention": { "label": "is sped up by …", "note": "Flash Attention speeds up the math without changing the result." } } }, "zh": { "fullName": "Self-Attention｜自注意力", "factExplain": "让序列中每个位置彼此关注的信息机制。", "humanExplain": "一句话进来后，它会让每个词都四处打量，像饭桌上默默观察全场的人，谁重要就多记谁。\n\n它帮助模型抓上下文关系，是 Transformer 理解句子的关键部件。", "humanExplainDisplay": "一句话进来后，\n它会让每个词都==四处打量==，\n像饭桌上默默观察全场的人，\n谁重要就==多记谁==。\n\n它帮助模型抓上下文关系，\n是 Transformer 理解句子的关键部件。", "relationsNarrative": "Attention\n它是注意力机制的一种，让序列内部自己彼此关联。\n\nTransformer\nTransformer 把它作为核心部件来建模上下文关系。\n\nKV cache\n生成时会缓存它用到的键和值，减少重复计算。\n\nFlash Attention\nFlash Attention 主要是在工程上加速它的计算过程。", "relations": { "attention": { "label": "属于…的一种", "note": "它是注意力机制里的核心形式。" }, "transformer": { "label": "构成…核心", "note": "Transformer 主要靠它处理上下文。" }, "kv-cache": { "label": "推理时配合…", "note": "生成时会缓存键值，避免重复算。" }, "flash-attention": { "label": "被…加速", "note": "后者是在不改结果下优化计算。" } } } } }, { "id": "self-supervised-learning", "name": "SSL", "layer": "L2", "era": "2018", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "masked-language-modeling" }, { "to": "language-modeling" }, { "to": "bert" } ], "track": "history", "i18n": { "en": { "fullName": "Self-supervised Learning", "factExplain": "A training method where data creates its own labels for the model.", "humanExplain": "Self-supervised learning is a worksheet with answers on the back. The AI guesses first, then flips to check.\n\nIt often gives a model its first big practice during pretraining. Later, it helps with language tasks and recognition.", "humanExplainDisplay": "Self-supervised learning is a ==worksheet with answers on the back==.\nThe AI ==guesses first==,\nthen flips to check.\n\nIt often gives a model\nits first big practice during pretraining.\nLater, it helps with language tasks\nand recognition.", "relationsNarrative": "Pretraining\nSelf-supervised learning is often used in pretraining, so the model learns broad patterns first.\n\nMLM\nMLM hides part of the text and asks the model to guess it.\n\nLM\nLM predicts the next word, so the text provides the training signal.\n\nBERT\nBERT used self-supervised pretraining, then adapted to understanding tasks.", "relations": { "pretraining": { "label": "is often used in …", "note": "It is a common way to pretrain large models." }, "masked-language-modeling": { "label": "includes …", "note": "MLM hides words and trains the model to guess them." }, "language-modeling": { "label": "supports … training", "note": "Next-word prediction is also self-supervised learning." }, "bert": { "label": "helped build …", "note": "BERT got its base skills from self-supervised pretraining." } } }, "zh": { "fullName": "自监督学习（Self-supervised Learning）", "factExplain": "用数据自身生成标签来训练模型的方法。", "humanExplain": "没人判卷也能学？这招像让 AI 自己出练习册、自己对答案，老师只负责把卷子搬来。\n\n常用于预训练先打底，再迁移到理解、生成和识别任务。", "humanExplainDisplay": "没人判卷也能学？\n这招像让 AI\n==自己出练习册==、\n自己==对答案==，\n老师只负责把卷子搬来。\n\n常用于预训练先打底，\n再迁移到理解、\n生成和识别任务。", "relationsNarrative": "Pretraining\n它常被用在预训练阶段，让模型先从海量数据里学通用模式。\n\nMasked-language-modeling\n遮住一部分内容再预测，是自监督里的经典训练任务。\n\nLanguage-modeling\n预测下一个词，本质上也是让数据自己提供训练信号。\n\nBert\nBERT 靠这套方法先做预训练，再适配下游理解任务。", "relations": { "pretraining": { "label": "常用于…阶段", "note": "它是大模型预训练的常见套路。" }, "masked-language-modeling": { "label": "包含典型任务", "note": "遮盖再预测，是它的经典做法。" }, "language-modeling": { "label": "支撑…训练", "note": "预测下一个词，也是自监督形式。" }, "bert": { "label": "催生代表模型", "note": "BERT 靠自监督预训练打下基础。" } } } } }, { "id": "semantic-network", "name": "Semantic Network", "layer": "L2", "era": "1968", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "ontology" }, { "to": "knowledge-graph" }, { "to": "symbolic-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Semantic Network", "factExplain": "A way to show ideas as dots, with labeled lines between them.", "humanExplain": "A semantic network is a family tree for ideas. Each line has a name tag, like “dog is an animal.”\n\nIt helps AI store facts and answer questions. The AI follows the lines to find an answer.", "humanExplainDisplay": "A semantic network is a ==family tree for ideas==.\nEach line has a ==name tag==,\nlike “dog is an animal.”\n\nIt helps AI store facts\nand answer questions.\nThe AI follows the lines\nto find an answer.", "relationsNarrative": "KR\nA semantic network is a classic early form of KR.\n\nOntology\nOntology is stricter about concept rules than a semantic network.\n\nKnowledge Graph\nKnowledge Graphs inherited the dot-and-line structure from semantic networks.\n\nSymbolic AI\nSymbolic AI often uses clear links for rule-based reasoning.", "relations": { "knowledge-representation": { "label": "is a kind of …", "note": "A semantic network is a classic early KR method." }, "ontology": { "label": "organizes ideas for …", "note": "Ontology uses stricter rules for concepts and links." }, "knowledge-graph": { "label": "inspired …", "note": "Knowledge Graphs kept the dot-and-line idea." }, "symbolic-ai": { "label": "supports reasoning in …", "note": "Symbolic AI needs clear links to reason with rules." } } }, "zh": { "fullName": "语义网络", "factExplain": "用节点和关系边表示概念及其语义联系。", "humanExplain": "语义网络像江湖关系谱：谁拜谁为师、谁跟谁结盟，画上线，AI 才认得清门派。\n\n用于知识表示和问答，让机器顺关系找答案。", "humanExplainDisplay": "语义网络像==江湖关系谱==：\n谁拜谁为师、谁跟谁结盟，\n画上线，\nAI 才==认得清门派==。\n\n用于知识表示和问答，\n让机器顺关系\n找答案。", "relationsNarrative": "Knowledge Representation\n语义网络是早期知识表示的经典形式。\n\nOntology\n本体比语义网络更强调严格概念约束。\n\nKnowledge Graph\n知识图谱继承了语义网络的点边结构。\n\nSymbolic AI\n符号 AI 常借清晰关系进行规则推理。", "relations": { "knowledge-representation": { "label": "属于…方法", "note": "它是早期知识表示的经典形式。" }, "ontology": { "label": "为…组织概念", "note": "本体常用更严格的关系约束。" }, "knowledge-graph": { "label": "启发…结构", "note": "知识图谱继承了点边表示思路。" }, "symbolic-ai": { "label": "服务于…推理", "note": "符号 AI 依赖清晰关系做推理。" } } } } }, { "id": "semi-supervised-learning", "name": "Semi-Supervised Learning", "layer": "L2", "era": "1998", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "unsupervised-learning" }, { "to": "self-supervised-learning" }, { "to": "representation-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Semi-Supervised Learning", "factExplain": "Training AI with a few labeled examples and many unlabeled ones.", "humanExplain": "Semi-supervised learning is like sorting a huge box of animal photos. You get ten sticky notes, then match the rest by fur and ears.\n\nUse it when labels cost a lot, but raw data is everywhere. It learns the pattern without tagging every example.", "humanExplainDisplay": "Semi-supervised learning is like sorting\n==a huge box of animal photos==.\nYou get ==ten sticky notes==,\nthen match the rest by fur and ears.\n\nUse it when labels cost a lot,\nbut raw data is everywhere.\nIt learns the pattern\nwithout tagging every example.", "relationsNarrative": "Supervised Learning\nSemi-supervised learning keeps the same goal, but uses fewer labels.\n\nUnsupervised Learning\nIt uses structure in unlabeled data, not just labeled examples.\n\nSSL\nIt often uses SSL first, then learns from a few labels.\n\nRepresentation Learning\nGood representations make the small label set go further.", "relations": { "supervised-learning": { "label": "uses fewer labels than …", "note": "It keeps the supervised goal, but needs fewer labels." }, "unsupervised-learning": { "label": "borrows unlabeled data from …", "note": "It finds patterns inside unlabeled data." }, "self-supervised-learning": { "label": "often pairs with …", "note": "A model may use SSL first, then learn from a few labels." }, "representation-learning": { "label": "depends on …", "note": "Good representations help a small label set go further." } } }, "zh": { "fullName": "半监督学习", "factExplain": "用少量标注加大量未标注数据一起训练的方法。", "humanExplain": "半监督学习像改试卷时只知道几份标准答案，老师看字迹套路和错题堆，也能把大多数卷子分对类。\n\n适合标注贵但原始数据很多时用。", "humanExplainDisplay": "半监督学习像改试卷时\n只知道几份==标准答案==，\n老师看字迹套路和错题堆，\n也能把大多数卷子==分对类==。\n\n适合标注贵但\n原始数据很多时用。", "relationsNarrative": "Supervised Learning\n它延续监督学习目标，但没那么依赖大规模标注。\n\nUnsupervised Learning\n它会借用无标签数据里的结构和分布信息。\n\nSelf-Supervised Learning\n它常和自监督连用，先学表示再用少量标签。\n\nRepresentation Learning\n学到好的特征表示，是它效果提升的关键。", "relations": { "supervised-learning": { "label": "补足…数据缺口", "note": "它在监督学习上减少对标注的依赖。" }, "unsupervised-learning": { "label": "借…用无标签数据", "note": "它会从未标注数据里挖结构。" }, "self-supervised-learning": { "label": "常与…配合", "note": "常先自监督预训练，再少量标注微调。" }, "representation-learning": { "label": "依赖…学特征", "note": "核心是先学到好表示再带动分类。" } } } } }, { "id": "sentiment-analysis", "name": "Sentiment Analysis", "layer": "L4", "era": "2000s", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "text-classification" }, { "to": "natural-language-processing" }, { "to": "affective-computing" }, { "to": "classification" } ], "track": "history", "i18n": { "en": { "fullName": "Sentiment Analysis", "factExplain": "An NLP task that decides the mood or attitude in text.", "humanExplain": "Sentiment analysis is the friend reading the group chat first. It spots “thanks!” versus “thanks...”\n\nIt labels text as positive, neutral, or negative. Teams use it on reviews and support messages.", "humanExplainDisplay": "Sentiment analysis is the friend\nreading the ==group chat== first.\nIt spots ==“thanks!” versus “thanks...”==\n\nIt labels text as positive, neutral, or negative.\nTeams use it on reviews and support messages.", "relationsNarrative": "Text Cls\nSentiment analysis is one kind of Text Cls.\n\nNLP\nIt is a common NLP task for understanding text.\n\nAffective Computing\nIt gives Affective Computing mood clues from text.\n\nClassification\nIt often turns mood into a Classification problem.", "relations": { "text-classification": { "label": "is a type of …", "note": "Sentiment analysis gives text an emotion label." }, "natural-language-processing": { "label": "is a common task in …", "note": "It helps machines read the attitude in words." }, "affective-computing": { "label": "supports …", "note": "It turns text mood into a signal computers can use." }, "classification": { "label": "uses …", "note": "It often uses classes like positive, neutral, and negative." } } }, "zh": { "fullName": "情感分析", "factExplain": "判断文本情绪倾向的 NLP 任务。", "humanExplain": "情感分析像给留言盖章：夸的盖\"好评\"，骂的盖\"差评\"，阴阳怪气也躲不过。\n\n用于舆情和客服评论，先分好中差再处理。", "humanExplainDisplay": "情感分析像给留言\n==盖章==：\n夸的盖\"好评\"，骂的盖\"差评\"，\n==阴阳怪气也躲不过==。\n\n用于舆情和客服评论，\n先分好中差再处理。", "relationsNarrative": "Text Classification\n情感分析通常是文本分类的一种特例。\n\nNatural Language Processing\n它是 NLP 里最常见的文本理解任务之一。\n\nAffective Computing\n它为情感计算提供文本侧的情绪线索。\n\nClassification\n它常把情绪判断转成分类问题来做。", "relations": { "text-classification": { "label": "属于…的一类", "note": "情感分析本质是给文本打情绪标签。" }, "natural-language-processing": { "label": "是…常见任务", "note": "它让机器读懂文字里的态度。" }, "affective-computing": { "label": "服务…目标", "note": "它把文本情绪变成可计算信号。" }, "classification": { "label": "借用…方法", "note": "常把正面、中性、负面当类别。" } } } } }, { "id": "seq2seq", "name": "Seq2Seq", "layer": "L3", "era": "2014", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "attention" }, { "to": "speech-to-text" }, { "to": "llm" } ], "track": "history", "i18n": { "en": { "fullName": "Sequence-to-Sequence Model", "factExplain": "A model pattern that turns one sequence into another sequence.", "humanExplain": "Seq2Seq is like a cafeteria tray swap. You slide in one neat row, and it slides back a new row.\n\nYou meet it in translation, summaries, and speech-to-text. It is a classic frame for generating output step by step.", "humanExplainDisplay": "Seq2Seq is like a ==cafeteria tray swap==.\nYou slide in one neat row,\nand it slides back ==a new row==.\n\nYou meet it in translation, summaries, and speech-to-text.\nIt is a classic frame for generating output step by step.", "relationsNarrative": "Transformer\nMany later Seq2Seq tasks use the Transformer architecture instead.\n\nAttention\nAttention helps Seq2Seq focus on the key parts of the input.\n\nSTT\nSTT can be seen as turning one kind of sequence into another.\n\nLLM\nLLMs follow the same generation path, but at a much larger scale.", "relations": { "transformer": { "label": "often built with …", "note": "Many later Seq2Seq jobs moved to Transformer models." }, "attention": { "label": "uses … to focus", "note": "Attention helps it look at the important input parts while writing." }, "speech-to-text": { "label": "can power …", "note": "STT can be seen as changing sound steps into word steps." }, "llm": { "label": "came before …", "note": "LLMs kept its step-by-step generation idea, but grew much bigger." } } }, "zh": { "fullName": "序列到序列模型", "factExplain": "把一个序列转换成另一个序列的建模方式。", "humanExplain": "输入来一串，输出换一串，它干的就是“听完这版，给我换个版本”，翻译腔十足。\n\n常用于翻译、摘要和语音转文字，是经典生成任务框架。", "humanExplainDisplay": "输入来一串，\n输出换一串，\n它干的就是\n“==听完这版==，\n给我==换个版本==”，\n翻译腔十足。\n\n常用于翻译、\n摘要和语音转文字，\n是经典生成任务框架。", "relationsNarrative": "Transformer\n后来很多这类任务，改由 Transformer 架构来实现。\n\nAttention\n注意力机制让它在生成时，能对准输入重点。\n\nSpeech-to-text\n语音转文字常可看成把一种序列变成另一种序列。\n\nLLM\n大语言模型延续了它的生成路线，只是规模更大。", "relations": { "transformer": { "label": "常用…来实现", "note": "后来很多这类任务都改用 Transformer。" }, "attention": { "label": "靠…看重点", "note": "注意力让它生成时更会抓重点。" }, "speech-to-text": { "label": "可用于…", "note": "语音转文字常可视作这类映射任务。" }, "llm": { "label": "是…的前辈", "note": "很多大模型继承了它的生成思路。" } } } } }, { "id": "sequence-labeling", "name": "Sequence Labeling", "layer": "L4", "era": "1980s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "natural-language-processing" }, { "to": "hidden-markov-model" }, { "to": "connectionist-temporal-classification" } ], "track": "history", "i18n": { "en": { "fullName": "Sequence Labeling", "factExplain": "An NLP task that assigns a label to each item in a sequence.", "humanExplain": "Sequence labeling is name-tag day for every word. Bob says “person.” Paris says “place.” And “the” gets a tiny shrug.\n\nNLP uses it to split text into words. It also marks word jobs, names, and places.", "humanExplainDisplay": "Sequence labeling is ==name-tag day==\nfor every word.\nBob says “person.”\nParis says “place.”\nAnd “the” gets ==a tiny shrug==.\n\nNLP uses it to split text into words.\nIt also marks word jobs,\nnames, and places.", "relationsNarrative": "Supervised Learning\nSequence labeling usually trains on sequences with correct labels already filled in.\n\nNLP\nIt is a basic NLP task for word splitting and name finding.\n\nHMM\nHMM was a classic early model for sequence labeling.\n\nCTC\nCTC helps when inputs and outputs do not line up neatly.", "relations": { "supervised-learning": { "label": "trains with …", "note": "It needs a correct label for each position." }, "natural-language-processing": { "label": "belongs to …", "note": "NLP uses it for word splitting and name finding." }, "hidden-markov-model": { "label": "was modeled with …", "note": "HMM was a classic early method for this task." }, "connectionist-temporal-classification": { "label": "aligns with …", "note": "CTC helps when inputs and labels do not line up neatly." } } }, "zh": { "fullName": "Sequence Labeling（序列标注）", "factExplain": "为序列中每个元素预测标签的任务。", "humanExplain": "序列标注像快递分拣贴面单：流水线上每个词，都得贴清身份，人名地名别贴错。\n\n它用于分词、词性和实体识别，把文本抠成结构化信息。", "humanExplainDisplay": "序列标注像快递分拣：\n给每个词==贴面单==，\n每个词都要==贴清身份==，\n人名地名别贴错。\n\n它用于分词、词性\n和实体识别，\n把文本抠成结构化信息。", "relationsNarrative": "Supervised Learning\n序列标注通常用已标好标签的序列来训练。\n\nNLP\n它是 NLP 里的基础任务，常用于分词和实体识别。\n\nHMM\nHMM 是早期序列标注的经典建模方法。\n\nCTC\nCTC 处理输入输出难对齐的序列标注。", "relations": { "supervised-learning": { "label": "用…训练", "note": "需要给每个位置提供标准标签。" }, "natural-language-processing": { "label": "属于…任务", "note": "分词、词性和实体识别都常用。" }, "hidden-markov-model": { "label": "曾用…建模", "note": "HMM 是经典序列标注方法。" }, "connectionist-temporal-classification": { "label": "借…对齐", "note": "CTC 适合语音等难对齐序列。" } } } } }, { "id": "sequence-modeling", "name": "Sequence Modeling", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "transformer" }, { "to": "language-modeling" }, { "to": "seq2seq" } ], "track": "history", "i18n": { "en": { "fullName": "Sequence Modeling", "factExplain": "A way to learn how earlier items affect later items in a sequence.", "humanExplain": "Sequence modeling is like reading a messy group chat. One “fine.” means nothing until you saw the drama above.\n\nIt learns how each part follows the last. You meet it in text, speech, and data over time.", "humanExplainDisplay": "Sequence modeling is like reading\n==a messy group chat==.\nOne “fine.” means nothing\nuntil you saw the ==drama above==.\n\nIt learns how each part\nfollows the last.\nYou meet it in text, speech,\nand data over time.", "relationsNarrative": "Token\nSequence modeling treats tokens as pieces in a set order.\n\nTransformer\nTransformer is a key way to do modern sequence modeling.\n\nLM\nLM is the most common real-world task for sequence modeling.\n\nSeq2Seq\nSeq2Seq extends sequence modeling from one sequence to another.", "relations": { "token": { "label": "strings together …", "note": "It treats tokens as ordered pieces." }, "transformer": { "label": "advanced by …", "note": "Transformers make long-range patterns easier to learn." }, "language-modeling": { "label": "supports …", "note": "LM is the most common use of sequence modeling." }, "seq2seq": { "label": "extends into …", "note": "Seq2Seq handles a sequence in and a sequence out." } } }, "zh": { "fullName": "序列建模", "factExplain": "学习序列中前后依赖关系的建模方法。", "humanExplain": "像追连续剧，不能只看这一集谁在吵，还得记得前面埋了啥线，后面反转才看得懂。\n\n它建模前后顺序关系，常用于文本、语音和时间序列。", "humanExplainDisplay": "像追连续剧，\n不能只看这一集谁在吵，\n还得记得前面埋了啥==线==，\n后面反转才看得==懂==。\n\n它建模前后顺序关系，\n常用于文本、\n语音和时间序列。", "relationsNarrative": "Token\n它把 token 视作有先后关系的连续序列。\n\nTransformer\nTransformer 是现代序列建模的关键实现方式。\n\nLanguage Modeling\n语言建模是序列建模最常见的落地任务。\n\nSeq2Seq\nSeq2Seq 把序列建模扩展到序列到序列生成。", "relations": { "token": { "label": "按…串起来", "note": "它把离散片段当成有顺序的数据。" }, "transformer": { "label": "被…大幅推进", "note": "Transformer 让长程依赖建模更高效。" }, "language-modeling": { "label": "支撑…任务", "note": "语言建模是它最典型的应用场景。" }, "seq2seq": { "label": "扩展到…框架", "note": "Seq2Seq 处理输入输出都成序列的问题。" } } } } }, { "id": "sgd", "name": "SGD", "layer": "L2", "era": "1991", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "pretraining" }, { "to": "adam" }, { "to": "deep-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Stochastic Gradient Descent", "factExplain": "An optimizer that updates model parameters using small batches of data.", "humanExplain": "SGD is cooking soup for a school cafeteria. You taste one spoon, tweak the salt, and hope lunch survives.\n\nIt trains models by updating parameters with small data batches. It is a basic tool for deep learning, pretraining, and fine-tuning.", "humanExplainDisplay": "SGD is cooking soup\nfor a ==school cafeteria==.\nYou taste ==one spoon==,\ntweak the salt,\nand hope lunch survives.\n\nIt trains models by updating parameters\nwith small data batches.\nIt is a basic tool\nfor deep learning, pretraining, and fine-tuning.", "relationsNarrative": "Parameter\nSGD keeps adjusting parameters so the model gets closer to the target.\n\nPretraining\nPretraining uses SGD-like methods to keep updating model weights.\n\nAdam\nAdam grew from SGD and learned to adjust step sizes.\n\nDeep Learning\nDeep Learning needs optimizers like SGD to make models learn.", "relations": { "parameter": { "label": "updates … values", "note": "SGD keeps changing parameters to reduce error." }, "pretraining": { "label": "is used in …", "note": "Pretraining often uses SGD-like methods to update model weights." }, "adam": { "label": "came before …", "note": "Adam adds smarter step sizes to the SGD idea." }, "deep-learning": { "label": "powers … training", "note": "Deep learning models use optimizers like SGD to learn." } } }, "zh": { "fullName": "随机梯度下降", "factExplain": "用小批数据反复更新参数的优化方法。", "humanExplain": "它像爬山时起大雾：看不见整座山，就先探脚下这一小步，哪边更往下，下一步就朝哪边挪。\n\n它是训练模型的基础优化法，常用于预训练、微调，也是很多优化器的起点。", "humanExplainDisplay": "它像爬山时起大雾：\n看不见整座山，\n就先探脚下==这一小步==，\n哪边更往下，\n下一步就朝==哪边挪==。\n\n它是训练模型的基础优化法，\n常用于预训练、微调，\n也是很多优化器的起点。", "relationsNarrative": "Parameter\nSGD 的工作就是反复调整参数，让模型输出更接近目标。\n\nPretraining\n预训练阶段要靠 SGD 这类优化方法持续更新模型权重。\n\nAdam\nAdam 是在 SGD 思路上发展的，更会自己调步子大小。\n\nDeep-learning\n深度学习模型能学会东西，离不开 SGD 这类优化器。", "relations": { "parameter": { "label": "更新…数值", "note": "它通过反复调整参数让误差变小。" }, "pretraining": { "label": "用于…训练", "note": "大模型预训练通常靠它这类方法更新。" }, "adam": { "label": "是…前辈", "note": "Adam 在它基础上加了自适应机制。" }, "deep-learning": { "label": "支撑…训练", "note": "深度学习模型普遍靠它完成优化。" } } } } }, { "id": "shadow-ai", "name": "Shadow AI", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "enterprise-ai-deployment" }, { "to": "ai-governance-framework" }, { "to": "data-privacy" }, { "to": "data-exfiltration" } ], "track": "today", "i18n": { "en": { "fullName": "Shadow AI", "factExplain": "Employees use unapproved AI tools at work and bypass company controls.", "humanExplain": "Shadow AI is like using a secret lunchbox microwave under your desk. Your burrito gets hot fast, but the office fire alarm has no clue.\n\nYou see it in office work. It can save time, but it can leak data and hide work from audits.", "humanExplainDisplay": "Shadow AI is like using a ==secret lunchbox microwave==\nunder your desk.\nYour burrito gets hot fast,\nbut the ==office fire alarm== has no clue.\n\nYou see it in office work.\nIt can save time,\nbut it can leak data\nand hide work from audits.", "relationsNarrative": "Enterprise AI Deployment\nShadow AI often means official tools are too slow, so workers find their own path.\n\nAI Governance\nAI Governance brings Shadow AI into the open with rules, approval, and audits.\n\nData-privacy\nUnapproved tools may send sensitive data outside company walls.\n\nExfiltration\nA quick file upload can become real data leaving the company.", "relations": { "enterprise-ai-deployment": { "label": "pushes … forward", "note": "Private tool use shows gaps in the official AI rollout." }, "ai-governance-framework": { "label": "needs …", "note": "Governance sets rules, approvals, and audit trails." }, "data-privacy": { "label": "puts … at risk", "note": "Sensitive data may be sent to outside tools." }, "data-exfiltration": { "label": "can cause …", "note": "Unauthorized uploads can turn into data leaving the company." } } }, "zh": { "fullName": "影子 AI", "factExplain": "员工绕过组织管控使用未批准 AI 工具的现象。", "humanExplain": "影子 AI 就像下班把公文交给野路子代跑，活快了，公司门禁也白装了。\n\n它常见于企业办公，提效同时带来数据外流和审计盲区。", "humanExplainDisplay": "影子 AI 就像下班\n把公文交给==野路子代跑==，\n活快了，\n公司==门禁也白装了==。\n\n它常见于企业办公，\n提效同时带来数据外流和审计盲区。", "relationsNarrative": "Enterprise AI Deployment\n影子 AI 常说明官方工具慢，员工先自己找路。\n\nAI Governance\n治理框架用规则、审批和审计把它拉回明处。\n\nData-privacy\n未批准工具可能把敏感资料送出公司边界。\n\nExfiltration\n员工随手上传文件，可能变成真实数据外流。", "relations": { "enterprise-ai-deployment": { "label": "倒逼…", "note": "员工私用工具暴露企业部署缺口。" }, "ai-governance-framework": { "label": "需要…", "note": "治理框架给使用边界和审计规则。" }, "data-privacy": { "label": "威胁…", "note": "敏感资料可能被传进外部工具。" }, "data-exfiltration": { "label": "可能造成…", "note": "未授权上传会变成数据外流。" } } } } }, { "id": "shakey-the-robot", "name": "Shakey the Robot", "layer": "L1", "era": "1969", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "strips" }, { "to": "automated-planning" }, { "to": "robotics" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Shakey the Robot", "factExplain": "The first mobile robot that could sense a room and plan its moves.", "humanExplain": "Shakey was like a nervous kid with a cafeteria tray. It stopped, stared, then picked a path.\n\nIt linked seeing, planning, and moving. In AI history, it became an early model for Agents.", "humanExplainDisplay": "Shakey was like a ==nervous kid==\nwith a cafeteria tray.\nIt stopped, stared,\nthen ==picked a path==.\n\nIt linked seeing, planning,\nand moving.\nIn AI history,\nit became an early model for Agents.", "relationsNarrative": "STRIPS\nSTRIPS was first designed for Shakey's action planning.\n\nPlanning\nShakey moved automated planning from paper into real rooms.\n\nRobotics\nShakey made robots more than movers. They could sense and decide.\n\nAgent\nShakey was an early model with sensing, planning, and acting.", "relations": { "strips": { "label": "gave rise to …", "note": "STRIPS was first built for Shakey's action planning." }, "automated-planning": { "label": "tested …", "note": "Shakey brought planning algorithms into real rooms." }, "robotics": { "label": "pushed … toward smarts", "note": "Shakey helped robots sense the world and make choices." }, "agent": { "label": "became an early model for …", "note": "Shakey already had a loop of sensing, planning, and acting." } } }, "zh": { "fullName": "Shakey 机器人", "factExplain": "首个能感知环境并规划行动的移动机器人。", "humanExplain": "Shakey 像扫地机祖师爷：走两步先发呆，却真能看屋子自己找路。\n\n它串起视觉、规划、行动，是智能体研究早期样板。", "humanExplainDisplay": "Shakey 像\n==扫地机祖师爷==：\n走两步先发呆，\n却真能自己找路。\n\n它串起视觉、规划、行动，\n是智能体研究早期样板。", "relationsNarrative": "STRIPS\nSTRIPS 最早就是为它的行动规划而设计。\n\nPlanning\n它把自动规划从纸面推向真实房间。\n\nRobotics\n它让机器人不只会动，还会感知和决策。\n\nAgent\n它是具备感知、规划、行动闭环的早期样板。", "relations": { "strips": { "label": "催生…", "note": "STRIPS 最早服务于它的行动规划。" }, "automated-planning": { "label": "验证…", "note": "它把规划算法带进真实房间。" }, "robotics": { "label": "推动…智能化", "note": "它让机器人开始会感知和决策。" }, "agent": { "label": "成为…早期样板", "note": "它已具备感知、规划、行动闭环。" } } } } }, { "id": "shap", "name": "SHAP", "layer": "L2", "era": "2017", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "explainable-ai" }, { "to": "feature-engineering" }, { "to": "xgboost" } ], "track": "history", "i18n": { "en": { "fullName": "SHapley Additive exPlanations", "factExplain": "A method that shows how each feature pushes an AI prediction up or down.", "humanExplain": "SHAP is like a pizza receipt after a group order. Pepperoni pushed the price up. A coupon pulled it down.\n\nIt explains risk and medical models. It shows which details raised or lowered the prediction.", "humanExplainDisplay": "SHAP is like a ==pizza receipt==\nafter a group order.\nPepperoni ==pushed the price up==.\nA coupon pulled it down.\n\nIt explains risk and medical models.\nIt shows which details raised\nor lowered the prediction.", "relationsNarrative": "XAI\nSHAP is a common XAI method for showing feature effects.\n\nFeature-engineering\nSHAP shows how each feature pushes a prediction up or down.\n\nXGBoost\nTree SHAP often explains XGBoost and other tree models.", "relations": { "explainable-ai": { "label": "belongs to … methods", "note": "SHAP is a common XAI method for showing feature effects." }, "feature-engineering": { "label": "measures … impact", "note": "It shows whether a feature pushes a prediction up or down." }, "xgboost": { "label": "explains … predictions", "note": "Tree SHAP often explains tree model results, like XGBoost." } } }, "zh": { "fullName": "SHapley Additive exPlanations，夏普利加性解释", "factExplain": "用夏普利值解释特征对预测贡献的方法。", "humanExplain": "SHAP 像火锅局AA小票：这锅为啥这么贵，牛肉加钱，青菜省钱，一项项摊开。\n\n用于解释风控、医疗模型，让人看清预测依据。", "humanExplainDisplay": "SHAP 像==火锅局AA小票==：\n这锅为啥这么贵，\n==牛肉加钱，青菜省钱==，\n一项项摊开。\n\n用于解释风控、医疗模型，\n让人看清，\n预测依据。", "relationsNarrative": "XAI\nSHAP 是 XAI 里常用的特征归因方法。\n\nFeature-engineering\n它能显示每个特征如何推高或压低预测。\n\nXGBoost\nTree SHAP 常用于解释 XGBoost 等树模型。", "relations": { "explainable-ai": { "label": "属于…方法", "note": "SHAP 是常用的特征归因解释法。" }, "feature-engineering": { "label": "衡量…影响", "note": "它能显示特征推高或压低预测。" }, "xgboost": { "label": "解释…预测", "note": "Tree SHAP 常用于解释树模型结果。" } } } } }, { "id": "shrdlu", "name": "SHRDLU", "layer": "L4", "era": "1970", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "symbolic-ai" }, { "to": "natural-language-understanding" }, { "to": "automated-planning" }, { "to": "eliza" } ], "track": "history", "i18n": { "en": { "fullName": "SHRDLU", "factExplain": "An early AI program that understood English commands and moved blocks in a tiny world.", "humanExplain": "SHRDLU was like a tiny LEGO foreman. It understood block orders, but off the LEGO table, it was lost.\n\nResearchers used it to study early language understanding. In its small world, it could chat, reason, and move blocks.", "humanExplainDisplay": "SHRDLU was like a ==tiny LEGO foreman==.\nIt understood block orders,\nbut off the ==LEGO table==,\nit was lost.\n\nResearchers used it to study early language understanding.\nIn its small world,\nit could chat, reason, and move blocks.", "relationsNarrative": "Symbolic AI\nSHRDLU was one of the early wow moments for Symbolic AI.\n\nNLU\nSHRDLU turned English commands into block actions.\n\nPlanning\nSHRDLU planned the steps after it understood a command.\n\nELIZA\nELIZA mostly chatted, but SHRDLU could act in a small world.", "relations": { "symbolic-ai": { "label": "was an early star of …", "note": "It used rules to connect words with block actions." }, "natural-language-understanding": { "label": "showed an early form of …", "note": "It turned English commands into block moves." }, "automated-planning": { "label": "used … to order moves", "note": "After understanding a command, it planned the block steps." }, "eliza": { "label": "did more than …", "note": "ELIZA chatted, but SHRDLU could act in its small world." } } }, "zh": { "fullName": "积木世界自然语言理解程序", "factExplain": "一个在积木世界中理解并执行英文指令的早期 AI 程序。", "humanExplain": "SHRDLU 像迷你工地包工头：积木往哪搬都听懂，一出工地就抓瞎。\n\n用于早期语言理解研究，证明小世界里，AI 能对话、推理、办事。", "humanExplainDisplay": "SHRDLU 像\n==迷你工地包工头==：\n积木往哪搬都听懂，\n==一出工地就抓瞎==。\n\n用于早期语言理解研究，\n证明小世界里，\nAI 能对话、推理、办事。", "relationsNarrative": "Symbolic AI\n它是符号 AI 早期最惊艳的演示之一。\n\nNatural-language-understanding\n它把英文指令理解成可执行的积木操作。\n\nPlanning\n它理解命令后，还要规划搬积木步骤。\n\nELIZA\nELIZA主要会聊天，它还能在小世界里行动。", "relations": { "symbolic-ai": { "label": "代表…的早期高光", "note": "它用符号规则连接语言与动作。" }, "natural-language-understanding": { "label": "展示…的早期形态", "note": "它能把英文指令落到积木操作。" }, "automated-planning": { "label": "依赖…安排动作", "note": "理解命令后还要规划搬积木步骤。" }, "eliza": { "label": "比…更会办事", "note": "ELIZA会聊，它能在小世界里行动。" } } } } }, { "id": "siamese-network", "name": "Siamese Network", "layer": "L3", "era": "1993", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "metric-learning" }, { "to": "contrastive-learning" }, { "to": "embedding" }, { "to": "face-recognition" } ], "track": "history", "i18n": { "en": { "fullName": "Siamese Network", "factExplain": "A neural network with twin branches sharing rules to compare two items.", "humanExplain": "It is like two identical bouncers at a school dance. Each checks one ID with the same rulebook. Then they compare notes and give one match score.\n\nYou meet it in face unlock. It also helps signature checks and image search. Its job is to judge if two things are alike.", "humanExplainDisplay": "It is like two ==identical bouncers==\nat a school dance.\nEach checks one ID\nwith the ==same rulebook==.\nThen they compare notes\nand give one ==match score==.\n\nYou meet it in face unlock.\nIt also helps signature checks and image search.\nIts job is to judge\nif two things are alike.", "relationsNarrative": "Metric Learning\nA Siamese Network is a common similarity model in Metric Learning.\n\nContrastive Learning\nContrastive Learning trains it with pairs, so matching items move closer.\n\nEmbedding\nIt turns both inputs into embeddings, then compares their distance.\n\nFace Recognition\nFace Recognition can use it to decide if two faces are the same person.", "relations": { "metric-learning": { "label": "serves …", "note": "It is a classic model for Metric Learning." }, "contrastive-learning": { "label": "often trains with …", "note": "Contrastive Learning teaches it to pull similar pairs closer." }, "embedding": { "label": "outputs … before comparing", "note": "It turns each item into an embedding, then measures distance." }, "face-recognition": { "label": "helps verify …", "note": "Face Recognition can use it to check if two faces match." } } }, "zh": { "fullName": "孪生网络", "factExplain": "用共享权重的双塔比较样本相似度的神经网络。", "humanExplain": "孪生网络像双胞胎门卫：一个看你、一个看证件照，同一套眼光，碰头只报一个相似分。\n\n用于人脸、签名和搜图，先各自编码，再量距离。", "humanExplainDisplay": "孪生网络像==双胞胎门卫==：\n一个看你、一个看证件照，\n同一套眼光，\n碰头只报一个==相似分==。\n\n用于人脸、签名和搜图，\n先各自编码，\n再量距离。", "relationsNarrative": "Metric Learning\n孪生网络是度量学习里常见的相似度模型。\n\nContrastive Learning\n对比学习常用成对样本训练孪生网络。\n\nEmbedding\n它把两个输入编码成向量，再比较距离。\n\nFace Recognition\n人脸验证可用它判断两张脸是否同一人。", "relations": { "metric-learning": { "label": "服务于…", "note": "它是度量学习的经典网络结构。" }, "contrastive-learning": { "label": "常用…训练", "note": "对比学习常教它拉近相似样本。" }, "embedding": { "label": "输出…再比较", "note": "它先把样本变成向量再算距离。" }, "face-recognition": { "label": "用于…验证", "note": "人脸验证常用它判断是否同一人。" } } } } }, { "id": "sift", "name": "SIFT", "layer": "L2", "era": "1999", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "feature-engineering" }, { "to": "ransac" }, { "to": "cnn" } ], "track": "history", "i18n": { "en": { "fullName": "Scale-Invariant Feature Transform", "factExplain": "An algorithm for finding image clues stable under zoom and rotation.", "humanExplain": "SIFT is like sticking tiny name tags on the corners of a photo. Turn the photo or stand farther away, and the tags still say, “Yep, same spot.”\n\nIt is used for photo stitching and object matching. It is old-school Computer Vision muscle.", "humanExplainDisplay": "SIFT is like sticking ==tiny name tags==\non the corners of a photo.\nTurn the photo or stand farther away,\nand the tags still say,\n==“Yep, same spot.”==\n\nIt is used for photo stitching\nand object matching.\nIt is old-school Computer Vision muscle.", "relationsNarrative": "Computer Vision\nSIFT was a classic tool for image matching in Computer Vision.\n\nFeature-engineering\nSIFT hand-builds comparable features from local texture.\n\nRANSAC\nAfter SIFT matches points, RANSAC often removes bad matches.\n\nCNN\nCNN later replaced many hand-built features with learned ones.", "relations": { "computer-vision": { "label": "powered classic …", "note": "SIFT was a classic tool for matching images." }, "feature-engineering": { "label": "hand-builds …", "note": "SIFT turns local texture into features you can compare." }, "ransac": { "label": "pairs with … to filter matches", "note": "RANSAC often removes bad match points after SIFT matching." }, "cnn": { "label": "was replaced by …", "note": "CNN features later took over many vision jobs." } } }, "zh": { "fullName": "尺度不变特征变换", "factExplain": "一种检测并描述尺度、旋转不变局部特征的算法。", "humanExplain": "SIFT给照片角落盖防伪章：换个姿势拍都不怕，拍远、拍歪、转个身，仍认得出同一处。\n\n它用于图像拼接和物体匹配，是传统视觉硬功夫。", "humanExplainDisplay": "SIFT给照片角落\n==盖防伪章==：\n换个姿势拍都不怕，\n拍远、拍歪、转个身，仍认得出==同一处==。\n\n它用于图像拼接和物体匹配，\n是传统视觉硬功夫。", "relationsNarrative": "Computer Vision\nSIFT 是传统视觉里做图像匹配的经典工具。\n\nFeature-engineering\n它把局部纹理手工变成可比较的特征。\n\nRANSAC\n特征匹配后，RANSAC 常负责剔除错配点。\n\nCNN\nCNN 后来用学习到的特征替代许多手工特征。", "relations": { "computer-vision": { "label": "支撑传统…", "note": "它曾是图像匹配的经典工具。" }, "feature-engineering": { "label": "手工提取…", "note": "SIFT把局部纹理变成可比特征。" }, "ransac": { "label": "配合…筛匹配", "note": "RANSAC常用来剔除误匹配点。" }, "cnn": { "label": "被…逐步替代", "note": "深度特征后来接过许多视觉任务。" } } } } }, { "id": "silicon-photonics", "name": "Silicon photonics", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "ai-chip" }, { "to": "ai-data-center" }, { "to": "memory-bandwidth" } ], "track": "today", "i18n": { "en": { "fullName": "Silicon photonics", "factExplain": "A technology that uses silicon chips to move data with light.", "humanExplain": "Silicon photonics gives chip data a tiny light-speed express lane. The copper wires sit in traffic and honk.\n\nIt helps chip links and AI data centers. It moves data faster and with less power.", "humanExplainDisplay": "Silicon photonics gives chip data\na tiny ==light-speed express lane==.\nThe ==copper wires== sit in traffic and honk.\n\nIt helps chip links and AI data centers.\nIt moves data faster and with less power.", "relationsNarrative": "AI chip\nSilicon photonics moves some chip-to-chip data from wires to light paths.\n\nAI data center\nAI data centers use it to raise link speed and cut power use.\n\nMemory bandwidth\nIt helps ease the old bottleneck of data moving slower than compute.", "relations": { "ai-chip": { "label": "speeds links for …", "note": "Silicon photonics can ease traffic between AI chips." }, "ai-data-center": { "label": "connects …", "note": "AI data centers use it to move data with less power." }, "memory-bandwidth": { "label": "eases … bottlenecks", "note": "Light links move data faster and use less power." } } }, "zh": { "fullName": "硅光子学", "factExplain": "用硅芯片操控光信号传输数据的技术。", "humanExplain": "硅光子像给芯片数据开地铁专线：电线堵成晚高峰，光路一路绿灯。\n\n它用于芯片互连和数据中心，降低传输瓶颈。", "humanExplainDisplay": "硅光子像给芯片数据\n开==地铁专线==：\n电线堵成晚高峰，\n光路一路绿灯。\n\n它用于芯片互连\n和数据中心，\n降低传输瓶颈。", "relationsNarrative": "AI Chip\n硅光子把芯片间通信从电线部分搬到光路。\n\nAI Data Center\nAI 数据中心用它提升互连带宽、降低功耗。\n\nMemory Bandwidth\n它缓解数据搬运慢过计算的老瓶颈。", "relations": { "ai-chip": { "label": "为…提速互连", "note": "硅光子可缓解芯片间通信瓶颈。" }, "ai-data-center": { "label": "连接…", "note": "数据中心用它降低传输功耗。" }, "memory-bandwidth": { "label": "缓解…瓶颈", "note": "光互连让数据搬运更快更省电。" } } } } }, { "id": "simclr", "name": "SimCLR", "layer": "L3", "era": "2020", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "contrastive-learning" }, { "to": "self-supervised-learning" }, { "to": "data-augmentation" }, { "to": "representation-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Simple Framework for Contrastive Learning of Visual Representations", "factExplain": "A self-supervised way to train image features by comparing picture pairs.", "humanExplain": "SimCLR is like a school-dance bouncer for photos. Same kid with sunglasses gets in. The look-alike waits outside.\n\nIt pretrains vision models on unlabeled images. Later, they need fewer labels to learn useful features.", "humanExplainDisplay": "SimCLR is like a ==school-dance bouncer== for photos.\nSame kid with sunglasses gets in.\nThe ==look-alike waits outside==.\n\nIt pretrains vision models on unlabeled images.\nLater, they need fewer labels\nto learn useful features.", "relationsNarrative": "Contrastive Learning\nSimCLR is a classic Contrastive Learning framework for images.\n\nSSL\nSimCLR uses unlabeled images for SSL pretraining.\n\nData Augmentation\nData Augmentation makes two views of the same image for SimCLR.\n\nRepresentation Learning\nSimCLR aims to learn visual features for Representation Learning.", "relations": { "contrastive-learning": { "label": "implements …", "note": "It pulls matching views together and pushes non-matches apart." }, "self-supervised-learning": { "label": "is a … method", "note": "It can pretrain without human labels." }, "data-augmentation": { "label": "uses … to make pairs", "note": "Two changed views of one image become a positive pair." }, "representation-learning": { "label": "learns … features", "note": "Its goal is image features that transfer to new tasks." } } }, "zh": { "fullName": "视觉表征对比学习简单框架", "factExplain": "用对比学习训练图像表征的自监督框架。", "humanExplain": "SimCLR 像相亲认照片：滤镜换角度也认本人，撞脸再像也不乱牵线。\n\n用于视觉预训练，少标注也能学到好特征。", "humanExplainDisplay": "SimCLR 像==相亲认照片==：\n滤镜换角度也认本人，\n撞脸再像，\n也==不乱牵线==。\n\n用于视觉预训练，\n少标注也能\n学到好特征。", "relationsNarrative": "Contrastive Learning\nSimCLR 是视觉对比学习的代表框架。\n\nSSL\n它用无标签图像做自监督预训练。\n\nData Augmentation\n图像增强为同一张图造出正样本对。\n\nRepresentation Learning\n最终目标是学到可迁移的视觉表征。", "relations": { "contrastive-learning": { "label": "实现…思路", "note": "把正样本拉近，负样本推远。" }, "self-supervised-learning": { "label": "属于…方法", "note": "不用人工标签也能预训练。" }, "data-augmentation": { "label": "依赖…造样本", "note": "同图两种增强视图构成正样本。" }, "representation-learning": { "label": "学习…特征", "note": "目标是学出可迁移图像表征。" } } } } }, { "id": "singularity", "name": "Singularity", "layer": "L6", "era": "1993", "publishedAt": "2026-05-23T11:55:00Z", "relations": [ { "to": "agi" }, { "to": "superintelligence" }, { "to": "alignment" } ], "i18n": { "en": { "fullName": "Singularity", "factExplain": "A hypothetical point where AI races beyond human understanding and control.", "humanExplain": "The Singularity is when the robot vacuum finds the turbo button. Cute one minute. Redecorating the house at 3 a.m. the next.\n\nIt is a sci-fi “what if,” not a sure event. It imagines super-smart AI racing past control, then changing society too fast to steer.", "humanExplainDisplay": "The Singularity is when the ==robot vacuum==\nfinds the ==turbo button==.\nCute one minute.\nRedecorating the house at 3 a.m. the next.\n\nIt is a sci-fi “what if,”\nnot a sure event.\nIt imagines super-smart AI racing past control,\nthen changing society too fast to steer.", "relationsNarrative": "AGI\nAGI is the usual starting point for Singularity talk.\n\nSuperintelligence\nSuperintelligence is the force that could make the Singularity speed up.\n\nAlignment\nAlignment decides whether Singularity risks can be held in check.", "relations": { "agi": { "label": "could follow from …", "note": "AGI is the usual starting point in Singularity talk." }, "superintelligence": { "label": "could be driven by …", "note": "Superintelligence is the key speed boost in the Singularity idea." }, "alignment": { "label": "makes … urgent", "note": "Alignment may help keep Singularity risks under control." } } }, "zh": { "fullName": "奇点", "factExplain": "AI 能力快速超越人类理解和控制范围的假想转折点。", "humanExplain": "奇点像打工人按电梯去上班，门一开却变火箭，直接冲出小区。\n\n它常用于讨论通用智能、超级智能和安全治理；未必会来，但值得提前想。", "humanExplainDisplay": "奇点像==打工人按电梯去上班==，\n门一开却==变火箭==，\n直接冲出小区。\n\n它常用于讨论通用智能、\n超级智能和安全治理；\n未必会来，\n但值得提前想。", "relationsNarrative": "AGI\nAGI 是 Singularity 讨论中最常见的前提条件。\n\nSuperintelligence\nSuperintelligence 是 Singularity 假设中的关键加速力量。\n\nAlignment\nAlignment 决定 Singularity 风险是否有可能被约束。", "relations": { "agi": { "label": "源于…加速" }, "superintelligence": { "label": "源于…失控" }, "alignment": { "label": "强化…的重要性" } } } } }, { "id": "situation-calculus", "name": "SitCalc", "layer": "L2", "era": "1963", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "frame-problem" }, { "to": "agent" }, { "to": "logic-programming" } ], "track": "history", "i18n": { "en": { "fullName": "Situation Calculus", "factExplain": "A logic form for describing actions, world states, and their results.", "humanExplain": "SitCalc sounds like math class, but relax. It is like a board game rule sheet. Move the robot, knock over the popcorn, and it tells you what changed.\n\nIt describes the world before and after an action. You see it in planning and robots. It also helps run pretend tasks first.", "humanExplainDisplay": "SitCalc sounds like math class,\nbut relax.\nIt is like a ==board game rule sheet==.\nMove the robot,\nknock over the popcorn,\nand it tells you ==what changed==.\n\nIt describes the world before and after an action.\nYou see it in planning and robots.\nIt also helps run pretend tasks first.", "relationsNarrative": "KR\nSituation Calculus is a classic KR method for actions and world changes.\n\nFrame Problem\nIt helps ask what changes after an action and what stays the same.\n\nAgent\nAn Agent can use it to plan action needs and results.\n\nLogic\nBoth use rules and clear reasoning.", "relations": { "knowledge-representation": { "label": "is a kind of …", "note": "SitCalc is a classic KR way to describe actions." }, "frame-problem": { "label": "helps handle …", "note": "It shows what changes after an action and what stays true." }, "agent": { "label": "helps … plan", "note": "An Agent can use it to state action needs and results." }, "logic-programming": { "label": "shares rules with …", "note": "Both use rules and clear reasoning." } } }, "zh": { "fullName": "Situation Calculus（情境演算）", "factExplain": "一种描述行动、状态变化与结果的逻辑形式。", "humanExplain": "情境演算像前台交接班记事本：谁刚办完入住、哪间房改了状态、下个班的人一翻就能接着干。\n\n它用来表示动作前后世界怎么变，常用于规划、机器人和任务推演。", "humanExplainDisplay": "情境演算像前台交接班记事本：\n谁刚办完入住，\n哪间房==改了状态==，\n下个班的人一翻\n就能==接着干==。\n\n它用来表示动作前后\n世界怎么变，\n常用于规划、机器人和任务推演。", "relationsNarrative": "Knowledge Representation\n它是知识表示中描述行动与世界变化的一种经典方法。\n\nFrame Problem\n它经常用来讨论动作发生后，哪些事实该变、哪些保持不变。\n\nAgent\nAgent 做任务规划时，可用它表示动作条件和执行结果。\n\nLogic Programming\n两者都属于符号主义方法，强调规则和显式推理。", "relations": { "knowledge-representation": { "label": "属于…方法", "note": "它是经典知识表示里的动作表示法。" }, "frame-problem": { "label": "用来处理…", "note": "它常被拿来表达动作后的不变与变化。" }, "agent": { "label": "帮助…做规划", "note": "Agent 可用它描述行动前提与结果。" }, "logic-programming": { "label": "同属符号路线", "note": "两者都强调规则、推理与显式表示。" } } } } }, { "id": "slam", "name": "SLAM", "layer": "L4", "era": "1986", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "kalman-filter" }, { "to": "computer-vision" }, { "to": "spatial-intelligence" } ], "track": "history", "i18n": { "en": { "fullName": "Simultaneous Localization and Mapping", "factExplain": "A technique that maps a place while tracking its own location.", "humanExplain": "SLAM is like your first trip through a giant IKEA. You look for the bathrooms while drawing the maze in your head.\n\nIt helps machines map new places and know their own spot. You meet it in robots, driverless cars, and AR apps.", "humanExplainDisplay": "SLAM is like your first trip\nthrough a ==giant IKEA==.\nYou look for the bathrooms\nwhile drawing ==the maze in your head==.\n\nIt helps machines map new places\nand know their own spot.\nYou meet it in robots,\ndriverless cars,\nand AR apps.", "relationsNarrative": "Robotics\nSLAM helps a robot know where it is while it moves.\n\nKalman Filter\nSLAM often uses a Kalman Filter to mix sensor data and estimate pose.\n\nComputer Vision\nVisual SLAM uses image features to rebuild the shape of a place.\n\nSpatial Intelligence\nSLAM gives Spatial Intelligence a map and a location to start from.", "relations": { "robotics": { "label": "supports … navigation", "note": "SLAM tells a robot where it is as it moves." }, "kalman-filter": { "label": "uses … to estimate state", "note": "A Kalman Filter helps smooth noisy sensor data." }, "computer-vision": { "label": "sees space with …", "note": "Visual SLAM uses images to understand space." }, "spatial-intelligence": { "label": "gives … a base map", "note": "Maps and location are the start of spatial intelligence." } } }, "zh": { "fullName": "同步定位与地图构建", "factExplain": "边建地图边估计自身位置的技术。", "humanExplain": "SLAM 像第一次逛大商场：人还没摸清方向，就边找厕所边在脑子里画地图。\n\n用于机器人、无人车和 AR，让机器在陌生地不迷路。", "humanExplainDisplay": "SLAM 像==第一次逛大商场==：\n人还没摸清方向，\n就边找厕所，\n边在脑子里==画地图==。\n\n用于机器人、无人车和 AR，\n让机器在陌生地，\n不迷路。", "relationsNarrative": "Robotics\nSLAM 让机器人边移动边知道自己在哪。\n\nKalman Filter\n它常用于融合传感器并估计位姿。\n\nComputer Vision\n视觉 SLAM 用图像特征还原环境结构。\n\nSpatial Intelligence\n它为机器理解三维空间提供位置和地图。", "relations": { "robotics": { "label": "支撑…导航", "note": "SLAM 让机器人知道自己在哪。" }, "kalman-filter": { "label": "常用…估计状态", "note": "滤波器帮助压住传感器噪声。" }, "computer-vision": { "label": "借助…看路", "note": "视觉 SLAM 依赖图像理解空间。" }, "spatial-intelligence": { "label": "提供…底图", "note": "地图和位置是空间智能的入口。" } } } } }, { "id": "small-language-model", "name": "SLM", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "llm" }, { "to": "local-llm" }, { "to": "quantization" }, { "to": "distillation" } ], "track": "today", "i18n": { "en": { "fullName": "Small Language Model", "factExplain": "A smaller language model that needs less memory and computing power.", "humanExplain": "An SLM is like a zippy scooter for quick errands. It will not move a couch, but it starts fast.\n\nIt often runs on your device or a small service. It fits daily tasks with quick replies and low cost.", "humanExplainDisplay": "An SLM is like a ==zippy scooter==\nfor quick errands.\nIt will not move a couch,\nbut it ==starts fast==.\n\nIt often runs on your device\nor a small service.\nIt fits daily tasks\nwith quick replies and low cost.", "relationsNarrative": "LLM\nAn SLM is the same kind of tool as an LLM, but lighter and cheaper to run.\n\nLocal-LLM\nAn SLM is easier to run locally because it needs less from the device.\n\nQuantization\nQuantization can shrink an SLM even more, so it is easier to run.\n\nDistillation\nDistillation can move skills from a large model into a smaller model.", "relations": { "llm": { "label": "is lighter than …", "note": "It is still a language model, but uses less compute and memory." }, "local-llm": { "label": "often powers …", "note": "Small models are easier to run on local devices." }, "quantization": { "label": "shrinks further with …", "note": "Quantization helps it fit on phones and computers." }, "distillation": { "label": "can be made by …", "note": "Distillation often turns a large model into a lighter one." } } }, "zh": { "fullName": "小语言模型（Small language model）", "factExplain": "参数规模较小、资源需求更低的语言模型。", "humanExplain": "别拿它跟大模型比块头，它更像楼下便利店：货品不全，但近、快、随到随取，应急特别顶用。\n\n常跑在本地设备或轻量服务里，适合低延迟、低成本的日常任务。", "humanExplainDisplay": "别拿它跟大模型比块头，\n它更像==楼下便利店==：\n货品不全，但近、快、随到随取，\n应急特别==顶用==。\n\n常跑在本地设备或轻量服务里，\n适合低延迟、低成本的日常任务。", "relationsNarrative": "LLM\n它和大语言模型同类，但更轻、更省资源。\n\nLocal-LLM\n小模型更适合本地部署，设备要求没那么高。\n\nQuantization\n量化能进一步压缩它，降低运行门槛。\n\nDistillation\n蒸馏常用来把大模型能力迁到小模型上。", "relations": { "llm": { "label": "相比…更轻", "note": "同属语言模型，但更省算力与内存。" }, "local-llm": { "label": "常作为…主力", "note": "小模型更容易在本地设备部署运行。" }, "quantization": { "label": "常配合…压缩", "note": "量化后更容易塞进手机和电脑。" }, "distillation": { "label": "可由…变小", "note": "常从大模型蒸馏出更轻版本。" } } } } }, { "id": "soft-actor-critic", "name": "SAC", "layer": "L2", "era": "2018", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "actor-critic" }, { "to": "off-policy-learning" }, { "to": "exploration-exploitation" }, { "to": "deep-reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Soft Actor-Critic", "factExplain": "An off-policy Deep RL method that rewards good actions and useful randomness.", "humanExplain": "SAC is like coaching a kid on a bike. You cheer smooth turns. You also cheer safe little experiments, not stiff mall-cop laps.\n\nIt is used in robots and simulations. It keeps control steady, while the AI still explores.", "humanExplainDisplay": "SAC is like coaching a ==kid on a bike==.\nYou cheer smooth turns.\nYou also cheer ==safe little experiments==,\nnot stiff mall-cop laps.\n\nIt is used in robots and simulations.\nIt keeps control steady,\nwhile the AI still explores.", "relationsNarrative": "Actor-Critic\nSAC adds an entropy reward to the Actor-Critic framework.\n\nOff-policy-learning\nSAC reuses past experience, so training can need fewer samples.\n\nExploration-Exploitation Tradeoff\nThe entropy reward leaves room to explore while still chasing rewards.\n\nDeep RL\nSAC is a common Deep RL method for continuous control.", "relations": { "actor-critic": { "label": "adds entropy to …", "note": "SAC adds an entropy reward to Actor-Critic." }, "off-policy-learning": { "label": "trains with …", "note": "It reuses old experience, so it needs fewer new samples." }, "exploration-exploitation": { "label": "eases …", "note": "The entropy reward keeps the policy from getting too rigid." }, "deep-reinforcement-learning": { "label": "belongs to …", "note": "SAC is a common Deep RL method for continuous control." } } }, "zh": { "fullName": "Soft Actor-Critic，软演员-评论家算法", "factExplain": "一种最大熵离策略深度强化学习算法。", "humanExplain": "SAC 像训狗捡球：不只奖捡回来，也奖多闻几条路，别只会直线冲。\n\n用于机器人和仿真控制，让策略稳，还愿意探索。", "humanExplainDisplay": "SAC 像训狗捡球：\n不只奖==捡回来==，\n也奖多闻几条路，\n别只会==直线冲==。\n\n用于机器人和仿真控制，\n让策略稳，\n还愿意探索。", "relationsNarrative": "Actor-Critic\nSAC 在演员-评论家框架上加入熵奖励。\n\nOff-policy Learning\nSAC 复用历史经验训练，样本效率更高。\n\nExploration-Exploitation\n熵奖励让策略在拿分外保留探索空间。\n\nDeep RL\nSAC 是连续控制里常用的深度强化学习算法。", "relations": { "actor-critic": { "label": "改造…框架", "note": "它在演员-评论家上加入熵奖励。" }, "off-policy-learning": { "label": "采用…训练", "note": "可复用旧经验，样本更省。" }, "exploration-exploitation": { "label": "缓解…取舍", "note": "熵奖励鼓励策略别太死板。" }, "deep-reinforcement-learning": { "label": "属于…", "note": "连续控制常用的深度强化学习算法。" } } } } }, { "id": "softmax", "name": "Softmax", "layer": "L2", "era": "1959", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "temperature" }, { "to": "llm" }, { "to": "backpropagation" } ], "track": "history", "i18n": { "en": { "fullName": "Softmax normalization function", "factExplain": "Turns a list of scores into probabilities with a total of 1.", "humanExplain": "Softmax is like cutting one pizza for a hungry table. The loudest kid gets the biggest slice.\n\nIt turns raw scores into fair-looking chances. You see it when AI picks a label or the next word.", "humanExplainDisplay": "Softmax is like cutting ==one pizza==\nfor a hungry table.\nThe ==loudest kid== gets the biggest slice.\n\nIt turns raw scores into fair-looking chances.\nYou see it when AI picks a label\nor the next word.", "relationsNarrative": "Token\nSoftmax turns each candidate Token score into a probability.\n\nTemperature\nTemperature can flatten or sharpen Softmax, so output feels more random or more focused.\n\nLLM\nAn LLM often uses Softmax to choose the next word.\n\nBackpropagation\nBackprop sends the error signal back through Softmax during training.", "relations": { "token": { "label": "assigns chances to …", "note": "Softmax turns candidate Token scores into probabilities." }, "temperature": { "label": "is shaped by …", "note": "Temperature makes its probabilities flatter or sharper." }, "llm": { "label": "helps … pick words", "note": "An LLM uses Softmax to turn scores into output probabilities." }, "backpropagation": { "label": "trains with …", "note": "Backprop sends the error signal back through Softmax." } } }, "zh": { "fullName": "Softmax 归一化函数", "factExplain": "把一组分数转成总和为 1 的概率。", "humanExplain": "像奶茶店分最后一桶珍珠：谁分得多、谁只捞几勺，得按人气折成一整桶里的占比。\n\n把一串分数压成概率，常用于分类和预测下一个词。", "humanExplainDisplay": "像奶茶店分最后一桶珍珠：\n谁分得==多==、谁只捞几勺，\n得按人气折成一整桶里的\n==占比==。\n\n把一串分数压成概率，\n常用于分类和预测下一个词。", "relationsNarrative": "Token\n它把每个候选词的分数转成可比较的概率。\n\nTemperature\n温度会拉平或拉尖它的分布，影响输出随机性。\n\nLLM\n大模型生成下一个词时，通常先靠它做概率化。\n\nBackpropagation\n训练时误差会穿过它回传，更新前面的参数。", "relations": { "token": { "label": "给…分配概率", "note": "它把候选词分数变成概率。" }, "temperature": { "label": "被…调节尖锐度", "note": "温度会改变概率分布的平滑程度。" }, "llm": { "label": "用于…选下一个词", "note": "大模型靠它把打分变成输出概率。" }, "backpropagation": { "label": "配合…训练", "note": "训练时梯度会穿过它往前传。" } } } } }, { "id": "software-engineering-benchmark", "name": "Software Engineering Benchmark", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-07-02T04:00:00.000Z", "relations": [ { "to": "agentic-coding" }, { "to": "model-leaderboard" }, { "to": "benchmark-contamination" }, { "to": "ai-qa-testing" } ], "track": "today", "i18n": { "en": { "fullName": "Software Engineering Benchmark", "factExplain": "A test for judging AI on real software engineering tasks.", "humanExplain": "It is a garage test for AI. The bike is broken. The wrench is missing. The chain still must work.\n\nIt compares coding assistants on real code projects. Can they read the project and fix a real bug?", "humanExplainDisplay": "It is a ==garage test== for AI.\nThe bike is broken.\nThe ==wrench is missing==.\nThe chain still must work.\n\nIt compares coding assistants\non real code projects.\nCan they read the project\nand fix a real bug?", "relationsNarrative": "Agentic coding\nThe benchmark checks if an agent can make real code changes.\n\nLeaderboard\nIts scores can go on a leaderboard for easy model comparison.\n\nBenchmark contamination\nIf tasks leak into training data, the score can look too high.\n\nAI QA Testing\nTest cases help prove the fix works and did not break the project.", "relations": { "agentic-coding": { "label": "tests … skills", "note": "It checks whether a coding agent can really change a project." }, "model-leaderboard": { "label": "feeds … rankings", "note": "Leaderboards use these scores to compare coding models." }, "benchmark-contamination": { "label": "can suffer from …", "note": "Leaked tasks can make the score look too high." }, "ai-qa-testing": { "label": "uses … to check fixes", "note": "Tests help show whether the bug fix really works." } } }, "zh": { "fullName": "软件工程基准测试", "factExplain": "用于评测 AI 解决真实软件工程任务的基准。", "humanExplain": "软件工程基准把 AI 拉进老小区修水管：图纸不全、管线乱，漏水真得堵上。\n\n它比较编程助手，能否读懂仓库并修好真 bug。", "humanExplainDisplay": "软件工程基准把 AI 拉进\n==老小区修水管==：\n图纸不全、管线乱，\n漏水==真得堵上==。\n\n它比较编程助手，\n能否读懂仓库\n并修好真 bug。", "relationsNarrative": "Agentic Coding\n软件工程基准常检验编码代理能否完成真实改动。\n\nLeaderboard\n它的分数常被整理成榜单，方便横向比较模型。\n\nBenchmark Contamination\n若题目进了训练数据，评测分数就可能虚高。\n\nAI QA Testing\n测试用例常作为判据，确认修复没有把项目改坏。", "relations": { "agentic-coding": { "label": "评测…能力", "note": "看编码代理能否真改项目。" }, "model-leaderboard": { "label": "支撑…排名", "note": "榜单常用它比较编程能力。" }, "benchmark-contamination": { "label": "警惕…污染", "note": "题目泄露会让分数虚高。" }, "ai-qa-testing": { "label": "借助…验证", "note": "测试用例常验证修复是否有效。" } } } } }, { "id": "sovereign-ai", "name": "Sovereign AI", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "on-premise-ai" }, { "to": "data-privacy" }, { "to": "ai-governance-framework" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "Sovereign AI", "factExplain": "AI systems and infrastructure controlled by a country or region itself.", "humanExplain": "Sovereign AI is a country keeping its own AI kitchen. The recipes stay home, and no delivery-app meltdown ruins dinner.\n\nIt keeps data local and puts local rules in charge. You meet it in public services, hospitals, and banks.", "humanExplainDisplay": "Sovereign AI is a country keeping its own ==AI kitchen==.\nThe ==recipes stay home==,\nand no delivery-app meltdown ruins dinner.\n\nIt keeps data local\nand puts local rules in charge.\nYou meet it in public services,\nhospitals, and banks.", "relationsNarrative": "On-premise AI\nSovereign AI often uses On-premise AI to keep data and systems under local control.\n\nData-privacy\nData-privacy is a main reason countries build Sovereign AI.\n\nAI Governance\nAI Governance sets the rules and limits for Sovereign AI.\n\nCompute-race\nSovereign AI projects make the Compute-race tougher.", "relations": { "on-premise-ai": { "label": "often runs as …", "note": "Sovereign AI often uses local deployment to keep control." }, "data-privacy": { "label": "is built around …", "note": "One key goal is keeping data inside the local boundary." }, "ai-governance-framework": { "label": "is bound by …", "note": "Governance rules decide how it is used and watched." }, "compute-race": { "label": "pushes the …", "note": "National AI buildouts raise demand for computing power." } } }, "zh": { "fullName": "主权 AI", "factExplain": "由国家或地区自主掌控的 AI 能力与基础设施。", "humanExplain": "这更像国家自己开的“电网级 AI”：数据不过境，规则自己写，外网抽风也不能停摆。\n\n多用于政务、医疗、金融，强调可控合规与本地运行。", "humanExplainDisplay": "这更像国家自己开的\n==电网级 AI==：\n==数据不过境==，规则自己写，\n外网抽风也不能停摆。\n\n多用于政务、医疗、金融，\n强调可控合规\n与本地运行。", "relationsNarrative": "On-premise AI\n主权 AI 常靠本地部署实现数据与系统可控。\n\nData-privacy\n数据隐私是主权 AI 最核心的建设动机之一。\n\nAI Governance\n治理框架决定主权 AI 的边界、责任与合规方式。\n\nCompute-race\n各国争建主权 AI，会进一步推高算力竞争。", "relations": { "on-premise-ai": { "label": "常用…落地", "note": "主权诉求常靠本地部署来实现。" }, "data-privacy": { "label": "围绕…建设", "note": "核心目标之一是守住数据边界。" }, "ai-governance-framework": { "label": "受…约束", "note": "治理框架决定它怎么管与怎么用。" }, "compute-race": { "label": "推动…升级", "note": "国家自建能力会加剧算力竞争。" } } } } }, { "id": "sparse-coding", "name": "Sparse Coding", "layer": "L2", "era": "1996", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "representation-learning" }, { "to": "unsupervised-learning" }, { "to": "autoencoder" }, { "to": "lasso" } ], "track": "history", "seo": { "zh": { "title": "稀疏编码 是什么?三个和弦弹流行歌,一文看懂 — AI Rookies", "description": "用少量基向量重构数据的表示学习方法。人话版一分钟讲明白。" }, "en": { "title": "What Is Sparse Coding? The Three Chord Singalong", "description": "A way to describe data using only a few learned building blocks. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Sparse Coding", "factExplain": "A way to describe data using only a few learned building blocks.", "humanExplain": "Sparse coding is like making a pop song with three guitar chords. Somehow, the crowd still sings along.\n\nModels use it to keep a few key patterns. It helps with features, noise cleanup, and compression.", "humanExplainDisplay": "Sparse coding is like making a ==pop song==\nwith ==three guitar chords==.\nSomehow,\nthe crowd still sings along.\n\nModels use it to keep\na few key patterns.\nIt helps with features,\nnoise cleanup,\nand compression.", "relationsNarrative": "Representation Learning\nSparse coding is Representation Learning with only a few key features.\n\nUnsupervised Learning\nSparse coding often learns its building blocks from unlabeled data.\n\nAutoencoder\nSparse autoencoders use the idea of keeping only a few neurons active.\n\nLasso\nLasso uses L1 pressure to force many numbers to zero.", "relations": { "representation-learning": { "label": "is a kind of …", "note": "It turns data into a few useful features." }, "unsupervised-learning": { "label": "often uses …", "note": "It can learn useful patterns without human labels." }, "autoencoder": { "label": "inspires sparse …", "note": "Sparse autoencoders use the same idea." }, "lasso": { "label": "uses …-style pressure", "note": "L1 pressure pushes many numbers to zero." } } }, "zh": { "fullName": "稀疏编码", "factExplain": "用少量基向量重构数据的表示学习方法。", "humanExplain": "稀疏编码像用三个和弦弹流行歌：和弦挑得少，靠搭配组合，整首歌照样还原。\n\n用于特征、降噪和压缩，让模型抓住少数关键模式。", "humanExplainDisplay": "稀疏编码像\n==用三个和弦弹流行歌==：\n和弦挑得少，靠搭配组合，\n==整首歌照样还原==。\n\n用于特征、降噪和压缩，\n让模型抓住少数关键模式。", "relationsNarrative": "Representation Learning\n稀疏编码是一种把数据变成少数关键特征的表示学习。\n\nUnsupervised Learning\n它常在无标签数据上学习字典和稀疏系数。\n\nAutoencoder\n稀疏自编码器把“少数神经元激活”用进网络。\n\nLasso\nLasso 的 L1 惩罚常用来逼出稀疏系数。", "relations": { "representation-learning": { "label": "学习稀疏…", "note": "它把数据压成少数有用特征。" }, "unsupervised-learning": { "label": "常用…训练", "note": "很多稀疏表示不用人工标签。" }, "autoencoder": { "label": "启发稀疏…", "note": "稀疏自编码器借了同一思路。" }, "lasso": { "label": "借助…约束", "note": "L1 惩罚会鼓励系数变少。" } } } } }, { "id": "spatial-intelligence", "name": "Spatial Intelligence", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "computer-vision" }, { "to": "embodied-ai" }, { "to": "world-model" }, { "to": "3d-ai-generation" } ], "track": "today", "i18n": { "en": { "fullName": "Spatial Intelligence", "factExplain": "The ability for AI to understand and reason about 3D space.", "humanExplain": "Spatial intelligence is the friend who can fit a couch through a doorway. It knows the hallway, the turn, and the magic tilt.\n\nYou meet it in robots and self-driving cars. It also helps 3D tools make scenes without wall-smashing weirdness.", "humanExplainDisplay": "Spatial intelligence is the friend\nwho can ==fit a couch through a doorway==.\nIt knows the hallway,\nthe turn,\nand the ==magic tilt==.\n\nYou meet it in robots and self-driving cars.\nIt also helps 3D tools make scenes\nwithout wall-smashing weirdness.", "relationsNarrative": "Computer Vision\nSpatial intelligence moves from seeing objects to understanding where they are.\n\nEmbodied AI\nEmbodied AI uses spatial intelligence to move, avoid obstacles, and handle objects.\n\nWorld model\nA world model needs spatial structure to predict real changes.\n\n3D AI Generation\n3D AI Generation uses spatial intelligence to make scenes feel more reasonable.", "relations": { "computer-vision": { "label": "adds space to …", "note": "It goes beyond seeing objects and understands where things are." }, "embodied-ai": { "label": "guides … action", "note": "A robot must understand space before it can move well." }, "world-model": { "label": "builds space for …", "note": "Space is a key part of modeling how the world changes." }, "3d-ai-generation": { "label": "makes … more believable", "note": "Spatial understanding helps 3D scenes avoid awkward mistakes." } } }, "zh": { "fullName": "Spatial Intelligence（空间智能）", "factExplain": "让 AI 理解并推理三维空间的能力。", "humanExplain": "空间智能像在宜家找出口：不只认沙发，还懂通道、拐角，和门朝哪开。\n\n用于机器人、自动驾驶和 3D 生成，减少撞墙穿帮。", "humanExplainDisplay": "空间智能像在\n==宜家找出口==：\n不只认沙发，\n还懂通道、拐角，和==门朝哪开==。\n\n用于机器人、自动驾驶\n和 3D 生成，\n减少撞墙穿帮。", "relationsNarrative": "Computer Vision\n空间智能把看见物体，推进到理解空间关系。\n\nEmbodied AI\n具身智能需要空间智能来导航、避障和操作。\n\nWorld Model\n世界模型要预测现实变化，离不开空间结构。\n\n3D AI Generation\n3D 生成有了空间智能，场景才更合理。", "relations": { "computer-vision": { "label": "扩展…到空间", "note": "不只识别图像，还理解位置关系。" }, "embodied-ai": { "label": "支撑…行动", "note": "机器人要先懂空间，才能动得对。" }, "world-model": { "label": "构建…的空间底座", "note": "空间理解是世界模型的重要部分。" }, "3d-ai-generation": { "label": "帮助…更合理", "note": "懂空间后，生成的 3D 更少穿帮。" } } } } }, { "id": "spec-to-code", "name": "Spec-to-code", "layer": "L5", "sublayer": "product", "era": "2020s", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "agentic-coding" }, { "to": "ai-app-builder" }, { "to": "structured-output" } ], "track": "today", "i18n": { "en": { "fullName": "Spec-to-code", "factExplain": "A process that turns written requirements into working code.", "humanExplain": "Spec-to-code is like ordering a custom sandwich at a busy deli. Say the bread and sauce, or you may get mayo on everything.\n\nIt turns written requirements into working code. Teams use it for quick prototypes and internal tools.", "humanExplainDisplay": "Spec-to-code is like ordering a ==custom sandwich==\nat a busy deli.\nSay the ==bread and sauce==,\nor you may get mayo on everything.\n\nIt turns written requirements\ninto working code.\nTeams use it for quick prototypes\nand internal tools.", "relationsNarrative": "Agentic coding\nClear specs give Agentic coding a clear goal and reduce rework.\n\nAI App Builder\nAn AI App Builder can turn specs into screens, data, and logic.\n\nStructured output\nStructured output writes requirements in a format the model can follow.", "relations": { "agentic-coding": { "label": "sets goals for …", "note": "Clear specs keep coding agents from wandering off." }, "ai-app-builder": { "label": "drives … to build apps", "note": "App builders often turn specs into screens and logic." }, "structured-output": { "label": "uses … for fixed format", "note": "Structured specs help models follow the request more reliably." } } }, "zh": { "fullName": "规格生成代码", "factExplain": "根据需求规格自动生成可运行代码的流程。", "humanExplain": "Spec-to-code 像网购定制柜：尺寸孔位写清楚，师傅才不装成鞋架。\n\n用于原型和内部工具，把需求快速变成可改代码。", "humanExplainDisplay": "Spec-to-code 像\n==网购定制柜==：\n尺寸孔位写清楚，\n师傅才==不装成鞋架==。\n\n用于原型和内部工具，\n把需求快速变成可改代码。", "relationsNarrative": "Agentic coding\n清晰规格给编码代理明确目标，减少来回返工。\n\nAI App Builder\n应用生成器常把规格直接变成页面、数据和逻辑。\n\nStructured output\n结构化输出能把需求写成模型更好执行的格式。", "relations": { "agentic-coding": { "label": "约束…的目标", "note": "清晰规格让编码代理少跑偏。" }, "ai-app-builder": { "label": "驱动…生成应用", "note": "应用生成器常从规格直接产出页面和逻辑。" }, "structured-output": { "label": "借…固定格式", "note": "结构化规格更容易被模型稳定执行。" } } } } }, { "id": "specialized-attention-hybridization", "name": "Specialized Attention Hybridization", "layer": "L3", "era": "2026", "publishedAt": "2026-07-10T05:00:00.000Z", "relations": [ { "to": "attention" }, { "to": "self-attention" }, { "to": "transformer" }, { "to": "flash-attention" } ], "track": "today", "seo": { "en": { "title": "What Is Specialized Attention Hybridization? Full and Linear Attention, Head by Head", "description": "Some attention heads need precision, others just need speed. A plain-English look at mixing full and linear attention inside one layer — and why it makes long-context models cheaper." }, "zh": { "title": "专化注意力混合是什么?按头分工的注意力混搭,一文看懂 — AI Rookies", "description": "同一层的注意力头各有特长:精细活给全注意力,巡场活给线性注意力。车间排班式的省算力思路,人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Specialized Attention Hybridization", "factExplain": "Mixing attention types per head: full attention for precise-recall heads, cheap linear attention for the rest.", "humanExplain": "Specialized attention hybridization is like staffing a workshop. Fine parts go to the master. Bulk parts go to the fast line.\n\nEach attention head gets the job it is best at. The model stays sharp and runs cheaper on long text.", "humanExplainDisplay": "Specialized attention hybridization\nis like ==staffing a workshop==.\nFine parts go to the master.\nBulk parts go to ==the fast line==.\n\nEach attention head gets\nthe job it is best at.\nThe model stays sharp\nand runs cheaper on long text.", "relationsNarrative": "Attention\nIt mixes full and linear attention inside one layer, head by head.\n\nSelf-Attention\nSelf-Attention heads in one layer do different jobs, so it matches tasks to strengths.\n\nTransformer\nIt makes Transformers sharp and cheap on long text.\n\nFlash Attention\nFlash Attention computes attention faster; this design computes less of it.", "relations": { "attention": { "label": "mixes … types", "note": "It mixes full and linear attention at head level." }, "self-attention": { "label": "splits … heads by role", "note": "Heads in one layer do different jobs, so tasks match strengths." }, "transformer": { "label": "speeds up …", "note": "It makes Transformers cheaper on long text." }, "flash-attention": { "label": "takes another road than …", "note": "Flash Attention computes faster; this computes less." } } }, "zh": { "fullName": "专化注意力混合", "factExplain": "按注意力头的功能特长混搭机制：精确召回的头用全注意力，其余头用低成本线性注意力。", "humanExplain": "专化注意力混合像车间排班：精密件交老师傅慢慢磨，大路货上流水线快速过，同一层里好钢用在刀刃上。\n\n用于长文本推理提效，精度不丢，算力省下。", "humanExplainDisplay": "专化注意力混合像\n==车间排班==：\n精密件交老师傅慢慢磨，\n大路货上流水线快速过，\n同一层里==好钢用在刀刃上==。\n\n用于长文本推理提效，\n精度不丢，算力省下。", "relationsNarrative": "Attention\n它在头一级混搭全注意力和线性注意力。\n\nSelf-Attention\n同一层的自注意力头功能各异，它按特长派活。\n\nTransformer\n它让 Transformer 跑长文本时又准又省。\n\nFlash Attention\nFlash Attention 把注意力算得更快，它让注意力算得更少。", "relations": { "attention": { "label": "混搭…机制", "note": "在头一级混用全注意力与线性注意力。" }, "self-attention": { "label": "按特长拆…的头", "note": "同层注意力头功能各异，按特长派活。" }, "transformer": { "label": "给…提效", "note": "让 Transformer 跑长文本时又准又省。" }, "flash-attention": { "label": "与…路线不同", "note": "Flash Attention 算得更快，它算得更少。" } } } } }, { "id": "spectral-clustering", "name": "Spectral Clustering", "layer": "L2", "era": "2001", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "clustering" }, { "to": "unsupervised-learning" }, { "to": "k-means-clustering" }, { "to": "dimensionality-reduction" } ], "track": "history", "i18n": { "en": { "fullName": "Spectral Clustering", "factExplain": "A method that groups items using eigenvectors from a similarity graph.", "humanExplain": "Spectral clustering is like sorting a messy school cafeteria. You ignore the table rows and watch who keeps trading fries.\n\nIt builds a similarity map, then finds groups in that map. It helps split images and spot circles in social networks.", "humanExplainDisplay": "Spectral clustering is like sorting\n==a messy school cafeteria==.\nYou ignore the table rows\nand watch who keeps ==trading fries==.\n\nIt builds a similarity map,\nthen finds groups in that map.\nIt helps split images\nand spot circles in social networks.", "relationsNarrative": "Clustering\nSpectral Clustering is a way to put similar items into groups.\n\nUnsupervised Learning\nSpectral Clustering usually groups data without human labels.\n\nK-Means Clustering\nAfter spectral embedding, K-Means Clustering often makes the final groups.\n\nDim. Reduction\nSpectral Clustering first turns the similarity graph into a smaller space.", "relations": { "clustering": { "label": "is a kind of …", "note": "It uses a similarity graph to decide who belongs together." }, "unsupervised-learning": { "label": "belongs to …", "note": "It usually groups data without human labels." }, "k-means-clustering": { "label": "often finishes with …", "note": "After spectral embedding, K-Means often makes the final groups." }, "dimensionality-reduction": { "label": "starts with …", "note": "It turns a similarity graph into a smaller space first." } } }, "zh": { "fullName": "谱聚类", "factExplain": "利用图谱特征向量进行聚类的方法。", "humanExplain": "谱聚类像广场舞分队：不看年龄职位，谁跟谁合拍就站一圈。\n\n擅长弯曲群体结构，用于图像分割和社交网络。", "humanExplainDisplay": "谱聚类像广场舞分队：\n不看==年龄职位==，\n谁跟谁合拍\n就==站一圈==。\n\n擅长弯曲群体结构，\n用于图像分割\n和社交网络。", "relationsNarrative": "Clustering\n谱聚类是把相似对象分成组的一种方法。\n\nUnsupervised Learning\n它通常不靠人工标签，而靠数据相似性分组。\n\nK-Means Clustering\n谱嵌入后，常用 K-Means 做最后分组。\n\nDimensionality Reduction\n它把相似图先变成更容易分组的低维表示。", "relations": { "clustering": { "label": "是一种…", "note": "用相似图来决定谁该分一组。" }, "unsupervised-learning": { "label": "属于…", "note": "通常不需要人工标签先带路。" }, "k-means-clustering": { "label": "常用…收尾", "note": "谱嵌入后再做最后分组。" }, "dimensionality-reduction": { "label": "先做…", "note": "把相似关系压成低维表示。" } } } } }, { "id": "speculative-decoding", "name": "Speculative Decoding", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "inference" }, { "to": "small-language-model" }, { "to": "llm" }, { "to": "tokens-per-second" } ], "track": "today", "i18n": { "en": { "fullName": "Speculative Decoding", "factExplain": "A speed trick: a small model drafts text, then a big model quickly checks it.", "humanExplain": "SpecDecode is like a drive-thru with a speedy rookie. The rookie lines up likely orders, and the manager just says yes or no.\n\nIt speeds up answers in online AI services. It makes words appear faster, not smarter.", "humanExplainDisplay": "SpecDecode is like a drive-thru\nwith a ==speedy rookie==.\nThe rookie lines up likely orders,\nand the manager just says ==yes or no==.\n\nIt speeds up answers\nin online AI services.\nIt makes words appear faster,\nnot smarter.", "relationsNarrative": "Inference\nSpecDecode happens during inference and aims to make text generation faster.\n\nSLM\nAn SLM often drafts the next tokens before the bigger model checks them.\n\nLLM\nSpecDecode reduces the LLM’s token-by-token work.\n\nTPS\nSpecDecode often shows up as higher TPS.", "relations": { "inference": { "label": "speeds up …", "note": "It speeds up decoding during inference." }, "small-language-model": { "label": "uses … to draft", "note": "An SLM often guesses the next text first." }, "llm": { "label": "takes work off …", "note": "The LLM checks more and generates less from scratch." }, "tokens-per-second": { "label": "raises …", "note": "One goal is higher text generation speed." } } }, "zh": { "fullName": "推测解码（Speculative decoding）", "factExplain": "用小模型预猜、多数结果由大模型快速验收的解码加速方法。", "humanExplain": "像烧烤摊先让小工把串排好，老板只负责扫一眼点头；多数时候不用每串都亲手重来。\n\n主要给大模型推理提速，常见于在线服务，提升出字速度，不直接增智。", "humanExplainDisplay": "像烧烤摊先让==小工把串排好==，\n老板只负责扫一眼==点头==；\n多数时候不用每串都亲手重来。\n\n主要给大模型推理提速，\n常见于在线服务，提升出字速度，\n不直接增智。", "relationsNarrative": "Inference\n它发生在推理阶段，核心目标是加快生成速度。\n\nSmall-language-model\n常先由小模型预猜 token，再交给大模型验证。\n\nLLM\n它给大模型减轻逐 token 生成负担。\n\nTokens-per-second\n推测解码常直接体现在 TPS 提升上。", "relations": { "inference": { "label": "用于加速…", "note": "它是推理阶段的解码提速手段。" }, "small-language-model": { "label": "常搭配…预猜", "note": "常先让小模型猜下一段内容。" }, "llm": { "label": "帮…减负", "note": "大模型更多负责验收而非全量生成。" }, "tokens-per-second": { "label": "提升…表现", "note": "目标之一就是提高生成速度指标。" } } } } }, { "id": "speech-recognition", "name": "ASR", "layer": "L4", "era": "1952", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "connectionist-temporal-classification" }, { "to": "natural-language-processing" }, { "to": "voice-to-voice-ai" } ], "track": "history", "i18n": { "en": { "fullName": "Automatic Speech Recognition", "factExplain": "A technology for turning spoken sound into text or basic speech sounds.", "humanExplain": "ASR is the drive-thru worker with one tired headset. Through wind and kid noise, it still hears “no pickles.”\n\nIt runs voice typing, live captions, and support call transcripts. One miss can push the next step off track.", "humanExplainDisplay": "ASR is the ==drive-thru worker==\nwith one tired headset.\nThrough wind and kid noise,\nit still hears ==“no pickles.”==\n\nIt runs voice typing,\nlive captions,\nand support call transcripts.\nOne miss can push the next step off track.", "relationsNarrative": "STT\nSTT is the most common output of ASR.\n\nCTC\nCTC often aligns flowing speech with text labels.\n\nNLP\nASR turns sound into text, then NLP tries to understand it.\n\nVoice-to-voice-ai\nASR is the ears of a voice chat system.", "relations": { "speech-to-text": { "label": "often outputs as …", "note": "STT is the most common way ASR shows up." }, "connectionist-temporal-classification": { "label": "aligns sound and text with …", "note": "CTC matches speech frames to shorter text labels." }, "natural-language-processing": { "label": "hands text to …", "note": "ASR turns speech into text, then NLP works on meaning." }, "voice-to-voice-ai": { "label": "serves as ears for …", "note": "Voice-to-voice-ai must hear the user before it can answer." } } }, "zh": { "fullName": "自动语音识别", "factExplain": "将语音信号识别为文字或音素的技术。", "humanExplain": "语音识别像烧烤摊老板听点单：风大口音杂，也得把羊肉串记成字。\n\n用于语音输入、字幕和客服转写，听错会带偏后续。", "humanExplainDisplay": "语音识别像烧烤摊老板\n==听点单==：\n风大口音杂，\n也得把羊肉串记成字。\n\n用于语音输入、字幕，\n和客服转写，\n听错会带偏后续。", "relationsNarrative": "STT\nSTT 是语音识别最常见的输出形式。\n\nCTC\nCTC 常用于对齐连续语音和文字标签。\n\nNLP\n语音识别把声音变文字，再交给 NLP 理解。\n\nVoice-to-voice AI\n语音识别是语音对话系统的耳朵。", "relations": { "speech-to-text": { "label": "常输出为…", "note": "STT 是语音识别最常见的落地形式。" }, "connectionist-temporal-classification": { "label": "用…对齐声音和文字", "note": "CTC 解决语音帧与字符长度不齐。" }, "natural-language-processing": { "label": "把文字交给…", "note": "识别后的文本还要继续被理解。" }, "voice-to-voice-ai": { "label": "作为…的耳朵", "note": "先听懂用户说什么，才能开口回答。" } } } } }, { "id": "speech-to-text", "name": "STT", "layer": "L4", "era": "2022", "publishedAt": "2026-05-31T00:57:33.087Z", "relations": [ { "to": "multimodal" }, { "to": "agent" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "Speech-to-Text", "factExplain": "Technology that turns spoken words into written text automatically.", "humanExplain": "STT is a meeting note-taker after three coffees. You are still talking, and the words race onto the screen.\n\nYou meet it in meeting notes, voice typing, and support call records. It saves your hands, but accents and noise can trip it up.", "humanExplainDisplay": "STT is a ==meeting note-taker==\nafter ==three coffees==.\nYou are still talking,\nand the words race onto the screen.\n\nYou meet it in meeting notes,\nvoice typing,\nand support call records.\nIt saves your hands,\nbut accents and noise can trip it up.", "relationsNarrative": "Multimodal AI\nSTT is a basic skill for Multimodal systems that handle speech input.\n\nAgent\nSTT lets an Agent receive tasks and instructions from speech.\n\nData-privacy\nSpeech transcripts often include personal details, so they can create privacy risks.", "relations": { "multimodal": { "label": "belongs to …", "note": "Speech understanding is a common entry point for Multimodal systems." }, "agent": { "label": "lets … hear speech", "note": "STT lets an Agent receive spoken instructions." }, "data-privacy": { "label": "can raise … issues", "note": "Speech transcripts can include private talks and personal details." } } }, "zh": { "fullName": "Speech-to-Text｜语音转文本", "factExplain": "把人类语音内容自动识别成文字的技术。", "humanExplain": "嘴还没停，字幕已经在前面狂奔，像旁边坐了个手速开挂的会议速记员。\n\n常用于会议转写、语音输入和客服记录；省手，但怕口音和杂音。", "humanExplainDisplay": "嘴还没停，\n字幕已经在前面==狂奔==，\n像旁边坐了个\n==手速开挂的会议速记员==。\n\n常用于会议转写、语音输入和客服记录；\n省手，但怕口音和杂音。", "relationsNarrative": "Multimodal AI\nSTT 是多模态系统处理语音输入的基础能力。\n\nAgent\nSTT 让 Agent 能从语音中接收任务和指令。\n\nData-privacy\n语音转写常包含个人信息，因此会带来隐私风险。", "relations": { "multimodal": { "label": "属于…能力", "note": "语音理解是多模态系统的常见入口。" }, "agent": { "label": "给…听懂人话", "note": "STT 让 Agent 能接收语音指令。" }, "data-privacy": { "label": "会碰到…问题", "note": "语音转写常涉及敏感对话与身份信息。" } } } } }, { "id": "stable-diffusion", "name": "Stable Diffusion", "layer": "L5", "sublayer": "product", "era": "2022", "publishedAt": "2026-07-05T04:00:00.000Z", "relations": [ { "to": "latent-diffusion-model" }, { "to": "diffusion" }, { "to": "text-to-image-generation" }, { "to": "clip" } ], "track": "history", "i18n": { "en": { "fullName": "Stable Diffusion", "factExplain": "A text-to-image model that uses latent diffusion to make pictures.", "humanExplain": "Stable Diffusion is like an Etch A Sketch with an art goblin inside. You type “pirate cat,” and TV static turns into a picture.\n\nIt can make a fresh image from words. It can also guide a sketch or fix a photo.", "humanExplainDisplay": "Stable Diffusion is like an ==Etch A Sketch==\nwith an ==art goblin== inside.\nYou type “pirate cat,”\nand TV static turns into a picture.\n\nIt can make a fresh image from words.\nIt can also guide a sketch\nor fix a photo.", "relationsNarrative": "LDM\nStable Diffusion is a famous LDM and puts diffusion in latent space.\n\nDiffusion\nDiffusion gives it the basic trick of turning noise into an image.\n\nText-to-Image Generation\nStable Diffusion made text-to-image tools much easier for regular people.\n\nCLIP\nCLIP helps it understand prompts and match them to image ideas.", "relations": { "latent-diffusion-model": { "label": "implements …", "note": "It is a famous LDM and does diffusion in latent space." }, "diffusion": { "label": "uses … denoising", "note": "Diffusion gives it the idea of turning noise into an image step by step." }, "text-to-image-generation": { "label": "popularized …", "note": "It helped regular people make images with plain text." }, "clip": { "label": "uses … for text", "note": "CLIP helps match the prompt to the image idea." } } }, "zh": { "fullName": "稳定扩散", "factExplain": "一种基于潜空间扩散的文生图模型。", "humanExplain": "稳定扩散像街头画像摊：你报长相，先糊一团，再越擦越像。\n\n用于文生图、草图和修图，文字就能出画。", "humanExplainDisplay": "稳定扩散像街头画像摊：\n你报==长相==，\n先糊一团，\n再==越擦越像==。\n\n用于文生图、草图和修图，\n文字就能出画。", "relationsNarrative": "Latent Diffusion Model\n它是 LDM 的代表实现，把扩散放进潜空间。\n\nDiffusion\nDiffusion 提供从噪声逐步还原图像的基本思路。\n\nText-to-Image Generation\n它把文生图带进大众视野，门槛大幅降低。\n\nCLIP\nCLIP 帮它理解提示词，并对齐图像语义。", "relations": { "latent-diffusion-model": { "label": "实现…", "note": "它是潜空间扩散模型的代表实现。" }, "diffusion": { "label": "沿用…去噪", "note": "扩散提供从噪声还原图像的思路。" }, "text-to-image-generation": { "label": "推动…普及", "note": "它让普通人用文字直接生成图片。" }, "clip": { "label": "借助…理解文本", "note": "CLIP 帮它把提示词对齐到图像。" } } } } }, { "id": "state-space-model", "name": "State Space Model", "layer": "L6", "era": "1960s", "publishedAt": "2026-06-21T14:27:14.376Z", "relations": [ { "to": "hidden-markov-model" }, { "to": "kalman-filter" }, { "to": "sequence-modeling" }, { "to": "latent-variable-model" } ], "track": "ingest", "i18n": { "en": { "fullName": "State Space Model (SSM)", "factExplain": "A model that tracks hidden state over time from noisy clues.", "humanExplain": "A state space model is like tracking a hamster inside your couch. You cannot see it, but the crumbs keep snitching.\n\nIt keeps a small hidden “state” and updates it step by step. You meet it in speech, sensors, finance, and newer long-context AI models.", "humanExplainDisplay": "A state space model is like tracking a hamster inside your couch.\nYou cannot see it, but the crumbs keep snitching.\n\nIt keeps a small hidden ==“state”== and updates it step by step.\nYou meet it in speech, sensors, finance, and newer long-context AI models.", "relations": { "hidden-markov-model": { "label": "is a parent idea of …", "note": "HMM is a simple state space model." }, "kalman-filter": { "label": "updates state with …", "note": "Kalman filters update hidden states." }, "sequence-modeling": { "label": "helps with …", "note": "SSMs handle streams one step at a time." }, "latent-variable-model": { "label": "uses …", "note": "The hidden state is a latent variable." } } } } }, { "id": "stateful-agent", "name": "Stateful agent", "layer": "L5", "sublayer": "architecture", "era": "2024", "publishedAt": "2026-05-28T15:58:23.418Z", "relations": [ { "to": "agent" }, { "to": "context-window" }, { "to": "rag" }, { "to": "function-call" } ], "track": "today", "i18n": { "en": { "fullName": "Stateful Agent", "factExplain": "An Agent that saves past state and uses it in later steps.", "humanExplain": "A Stateful Agent is the teammate who does not forget after lunch. It keeps the sticky note and starts on the right page.\n\nYou meet it in long chats and multi-step jobs. It can keep going, but its notes must stay tidy.", "humanExplainDisplay": "A Stateful Agent is the ==teammate who does not forget== after lunch.\nIt keeps the ==sticky note==\nand starts on the right page.\n\nYou meet it in long chats\nand multi-step jobs.\nIt can keep going,\nbut its notes must stay tidy.", "relationsNarrative": "Agent\nA Stateful Agent is an Agent that remembers history and progress.\n\nContext-window\nIt can save some information in state, so the Context-window feels less crowded.\n\nRAG\nWhen its own state is missing facts, RAG can fetch them.\n\nFunction-calling\nAfter it remembers progress, it can use Function-call across many steps.", "relations": { "agent": { "label": "is a kind of …", "note": "It is an Agent with memory of past work." }, "context-window": { "label": "eases … limits", "note": "Saved state can reduce pressure on the Context-window." }, "rag": { "label": "fills memory with …", "note": "RAG can bring back missing outside information." }, "function-call": { "label": "keeps using …", "note": "Remembered progress helps it call tools across many steps." } } }, "zh": { "fullName": "有状态代理", "factExplain": "一种能保存并利用历史状态的 Agent 形态。", "humanExplain": "它像认工位的老同事，昨天聊到哪、表填到哪，今天接着干。\n\n适合长任务、工作流和个人助理，能少重复交代，也更要管好记忆权限。", "humanExplainDisplay": "它像==认工位的老同事==，\n昨天聊到哪、表填到哪，\n==今天接着干==。\n\n适合长任务、工作流和个人助理，\n能少重复交代，\n也更要管好记忆权限。", "relationsNarrative": "Agent\nstateful agent 本质上是能记住历史与进度的 Agent。\n\nContext-window\n它可把部分信息留在状态里，减轻上下文窗口不够用的问题。\n\nRAG\n当自身状态不全时，RAG 可帮它补回需要的外部信息。\n\nFunction-calling\n记住任务进度后，它更适合连续调用工具完成多步操作。", "relations": { "agent": { "label": "是…的一种形态", "note": "它本质上是带记忆能力的 Agent。" }, "context-window": { "label": "缓解…限制", "note": "状态可部分减轻上下文装不下的问题。" }, "rag": { "label": "可用…补记忆", "note": "外部检索能帮它找回遗漏的信息。" }, "function-call": { "label": "配合…执行任务", "note": "记住进度后更适合连续调用工具。" } } } } }, { "id": "statistical-learning-theory", "name": "SLT", "layer": "L2", "era": "1960s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "bias-variance-tradeoff" }, { "to": "empirical-risk-minimization" }, { "to": "regularization" }, { "to": "scaling-law" } ], "track": "history", "i18n": { "en": { "fullName": "Statistical Learning Theory", "factExplain": "A theory about how models learn from samples and still work on new cases.", "humanExplain": "SLT is the coach who ignores your lucky half-court shot. It asks, can you score again on Saturday?\n\nIt helps explain generalization, overfitting, and sample size. You meet it in model training and testing.", "humanExplainDisplay": "SLT is the coach\nwho ignores your ==lucky half-court shot==.\nIt asks,\ncan you ==score again== on Saturday?\n\nIt helps explain generalization,\noverfitting,\nand sample size.\nYou meet it in model training\nand testing.", "relationsNarrative": "Bias-Variance Tradeoff\nSLT explains the cost when a model is too simple or too complex.\n\nERM\nERM is a key base for Statistical Learning Theory.\n\nRegularization\nRegularization can reduce overfitting and help models work on new data.\n\nScaling-law\nBoth study how data size links to model performance.", "relations": { "bias-variance-tradeoff": { "label": "explains …", "note": "It shows why too simple or too complex can hurt new data." }, "empirical-risk-minimization": { "label": "backs … with theory", "note": "ERM is a key starting point for SLT." }, "regularization": { "label": "shows why … helps", "note": "Regularization helps stop overfitting and improve new-data results." }, "scaling-law": { "label": "helps explain …", "note": "Both ask how data size changes model performance." } } }, "zh": { "fullName": "统计学习理论", "factExplain": "研究模型如何从样本中学会泛化的理论框架。", "humanExplain": "相亲不能光看精修照聊得嗨，关键是见面后靠不靠谱；它研究的就是这种“别被样本骗了”。\n\n它用来理解泛化、过拟合和样本规模，影响模型训练与评估。", "humanExplainDisplay": "相亲不能光看==精修照==聊得嗨，\n关键是见面后靠不靠谱；\n它研究的就是这种\n==“别被样本骗了”==。\n\n它用来理解泛化、过拟合和样本规模，\n影响模型训练与评估。", "relationsNarrative": "Bias-variance-tradeoff\n它解释模型太简单或太复杂时的泛化代价。\n\nEmpirical-risk-minimization\n经验风险最小化是统计学习理论的重要基础。\n\nRegularization\n正则化常被用来限制过拟合、改善泛化表现。\n\nScaling-law\n两者都在讨论数据规模与模型表现的关系。", "relations": { "bias-variance-tradeoff": { "label": "解释…取舍", "note": "它研究拟合能力与泛化误差平衡。" }, "empirical-risk-minimization": { "label": "给…打理论底", "note": "经验风险最小化是它的重要出发点。" }, "regularization": { "label": "说明…为何有用", "note": "正则化常被用来控制过拟合。" }, "scaling-law": { "label": "帮助理解…", "note": "都关心数据规模与性能怎么变化。" } } } } }, { "id": "statistical-machine-translation", "name": "SMT", "layer": "L4", "era": "1990", "publishedAt": "2026-06-29T04:00:00.000Z", "relations": [ { "to": "machine-translation" }, { "to": "neural-machine-translation" }, { "to": "n-gram-language-model" } ], "track": "history", "i18n": { "en": { "fullName": "Statistical Machine Translation", "factExplain": "A translation method that learns likely word matches from bilingual text.", "humanExplain": "SMT is like a tired subtitler with a giant stack of old DVDs. It sees “I’m fine” and bets on the line it has seen most.\n\nIt powered early web and document translation. It needed lots of matched bilingual text, then NMT mostly took over.", "humanExplainDisplay": "SMT is like a ==tired subtitler==\nwith a ==giant stack of old DVDs==.\nIt sees “I’m fine”\nand bets on the line it has seen most.\n\nIt powered early web and document translation.\nIt needed lots of matched bilingual text,\nthen NMT mostly took over.", "relationsNarrative": "MT\nSMT was the main MT route before neural networks.\n\nNMT\nNMT used one neural network instead of the old SMT setup.\n\nN-gram LM\nN-gram LM often gave each possible translation a smoothness score.\n\nIBM Models\nIBM Models laid the base for word alignment and translation odds.", "relations": { "machine-translation": { "label": "powered early …", "note": "SMT was the main MT path before neural models." }, "neural-machine-translation": { "label": "was replaced by …", "note": "NMT later became the main approach." }, "n-gram-language-model": { "label": "scored with …", "note": "N-gram LM helped judge if a translation sounded natural." } } }, "zh": { "fullName": "Statistical Machine Translation，统计机器翻译", "factExplain": "用统计模型从双语语料中学习翻译概率。", "humanExplain": "统计机翻像字幕组翻旧剧：台词库越厚，越会押最常见译法。\n\n曾支撑网页和文档翻译，依赖双语语料，后来多被神经机翻接棒。", "humanExplainDisplay": "统计机翻像\n==字幕组翻旧剧==：\n台词库越厚，\n越会押==最常见译法==。\n\n曾支撑网页和文档翻译，\n依赖双语语料，\n后来多被神经机翻接棒。", "relationsNarrative": "MT\nSMT 是机器翻译在神经网络前的主流路线。\n\nNMT\nNMT 用端到端神经网络取代统计机翻框架。\n\nN-gram LM\nN-gram LM 常用来给候选译文打顺口分。\n\nIBM Models\nIBM Models 奠定了词对齐和翻译概率基础。", "relations": { "machine-translation": { "label": "支撑早期…", "note": "统计机翻曾是机器翻译主流路线。" }, "neural-machine-translation": { "label": "被…接棒", "note": "神经机翻后来取代统计机翻主流。" }, "n-gram-language-model": { "label": "常用…评分", "note": "N-gram LM 帮它判断译文是否顺口。" } } } } }, { "id": "streaming-multimodal-model", "name": "Live Multimodal", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "voice-to-voice-ai" }, { "to": "inference" }, { "to": "tokens-per-second" } ], "track": "today", "i18n": { "en": { "fullName": "Streaming Multimodal Model", "factExplain": "A model that answers while live text, sound, images, or video come in.", "humanExplain": "It is like a drive-thru worker at lunch rush. You talk, point, and change your order, but they still keep up.\n\nIt answers while sound, video, or your screen comes in. Low delay makes voice chat, live translation, and screen help feel smooth.", "humanExplainDisplay": "It is like a ==drive-thru worker==\nat lunch rush.\nYou talk, point, and change your order,\nbut they ==still keep up==.\n\nIt answers while sound, video,\nor your screen comes in.\nLow delay makes voice chat,\nlive translation,\nand screen help feel smooth.", "relationsNarrative": "Multimodal AI\nA Streaming Multimodal Model turns many input types into a live stream.\n\nVoice-to-voice-ai\nVoice-to-voice-ai uses it to listen and answer at the same time.\n\nInference\nIt needs inference to keep taking input and sending output.\n\nTPS\nHigher TPS usually makes the reply feel more live.", "relations": { "multimodal": { "label": "extends … to live streams", "note": "Multimodal AI is the base for handling many input types." }, "voice-to-voice-ai": { "label": "powers real-time …", "note": "Listening and answering at once makes voice talk feel natural." }, "inference": { "label": "needs low-delay …", "note": "Streaming works when inference keeps sending small outputs." }, "tokens-per-second": { "label": "feels faster with higher …", "note": "More tokens per second makes live replies feel quicker." } } }, "zh": { "fullName": "流式多模态模型", "factExplain": "能边接收多模态输入边生成响应的模型。", "humanExplain": "流式多模态模型像篮球解说：球还在飞，就边看动作边接话。\n\n让语音助手、实时翻译、屏幕协助更顺，核心是低延迟。", "humanExplainDisplay": "流式多模态模型像\n==篮球解说==：\n球还在飞，\n就边看动作边接话。\n\n让语音助手、实时翻译、\n屏幕协助更顺，\n核心是低延迟。", "relationsNarrative": "Multimodal AI\n流式多模态模型把多种输入变成实时流。\n\nVoice-to-voice AI\n实时语音对话常依赖它边听边答。\n\nInference\n它要求推理边接收输入边持续输出。\n\nTPS\n输出速度越高，实时感通常越强。", "relations": { "multimodal": { "label": "扩展…到实时流", "note": "多模态是它处理多种输入的基础。" }, "voice-to-voice-ai": { "label": "支撑…实时对话", "note": "边听边答让语音对话更自然。" }, "inference": { "label": "要求…低延迟", "note": "流式体验靠推理过程不断产出。" }, "tokens-per-second": { "label": "受…影响体验", "note": "生成速度直接影响实时感。" } } } } }, { "id": "strips", "name": "STRIPS", "layer": "L2", "era": "1971", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "automated-planning" }, { "to": "symbolic-ai" }, { "to": "frame-problem" }, { "to": "situation-calculus" } ], "track": "history", "i18n": { "en": { "fullName": "Stanford Research Institute Problem Solver", "factExplain": "A classic planning format for action needs and action results.", "humanExplain": "STRIPS is like a school lunch ticket. It says, “Need bread first.” Then poof, bread down, sandwich up.\n\nIt turns a goal into small actions with needs and results. You meet it in old-school AI planning.", "humanExplainDisplay": "STRIPS is like a ==school lunch ticket==.\nIt says, “Need bread first.”\nThen poof,\n==bread down, sandwich up==.\n\nIt turns a goal into small actions\nwith needs and results.\nYou meet it in old-school AI planning.", "relationsNarrative": "Planning\nSTRIPS defines actions for classic planning with needs and results.\n\nSymbolic AI\nSTRIPS writes world states and actions as symbols.\n\nFrame Problem\nSTRIPS only lists changed facts and leaves the rest alone.\n\nSitCalc\nSTRIPS makes SitCalc-style action logic easier to run.", "relations": { "automated-planning": { "label": "defines actions for …", "note": "It writes each action with needs and results." }, "symbolic-ai": { "label": "borrows from …", "note": "It describes states and actions with symbols." }, "frame-problem": { "label": "simplifies …", "note": "It lists only the facts an action changes." }, "situation-calculus": { "label": "makes … more practical", "note": "It makes action logic easier to run." } } }, "zh": { "fullName": "Stanford Research Institute Problem Solver，斯坦福研究所问题求解器", "factExplain": "用前提和效果描述动作的经典规划形式。", "humanExplain": "把STRIPS想成煎饼摊接单：有蛋有肠才开做，做完少一份料、多一个成品。\n\n它把目标拆成带前提和效果的动作，方便规划搜索。", "humanExplainDisplay": "把STRIPS想成\n==煎饼摊接单==：\n有蛋有肠才开做，\n做完==少一份料==、多一个成品。\n\n它把目标拆成\n带前提和效果的动作，\n方便规划搜索。", "relationsNarrative": "Planning\nSTRIPS 用前提和效果，为经典规划定义动作。\n\nSymbolic AI\nSTRIPS 把世界状态和动作都写成符号。\n\nFrame Problem\nSTRIPS 只列变化项，简化“什么不变”的难题。\n\nSituation Calculus\nSTRIPS 是更工程化的动作逻辑表示。", "relations": { "automated-planning": { "label": "给…定义动作", "note": "用前提和效果把动作写清楚。" }, "symbolic-ai": { "label": "继承…思路", "note": "状态和动作都用符号描述。" }, "frame-problem": { "label": "简化…", "note": "只声明会改变的事实。" }, "situation-calculus": { "label": "工程化…", "note": "把动作逻辑写得更可执行。" } } } } }, { "id": "structural-causal-model", "name": "SCM", "layer": "L2", "era": "1995", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "do-calculus" }, { "to": "knowledge-representation" }, { "to": "world-model" }, { "to": "ai-bias" } ], "track": "history", "i18n": { "en": { "fullName": "Structural Causal Model", "factExplain": "A framework for mapping variables and their cause-and-effect links.", "humanExplain": "A Structural Causal Model is a detective board for cafeteria chaos. Did the spilled milk start the food fight, or just get blamed?\n\nIt helps separate real causes from tag-along patterns. You meet it in research and impact studies.", "humanExplainDisplay": "A Structural Causal Model is a ==detective board==\nfor cafeteria chaos.\nDid the ==spilled milk== start the food fight,\nor just get blamed?\n\nIt helps separate real causes\nfrom tag-along patterns.\nYou meet it in research\nand impact studies.", "relationsNarrative": "Do-Calculus\nA Structural Causal Model gives Do-Calculus a causal map for intervention reasoning.\n\nKR\nA Structural Causal Model is a KR method for showing cause and effect clearly.\n\nWorld model\nA World model can use it to organize states and causal effects.\n\nAI-bias\nIt helps show whether AI-bias comes from real causes or surface patterns.", "relations": { "do-calculus": { "label": "supports … interventions", "note": "It gives Do-Calculus the causal map it needs." }, "knowledge-representation": { "label": "is a kind of …", "note": "It writes cause-and-effect links into the knowledge." }, "world-model": { "label": "can shape a …", "note": "A causal World model can use it to organize variables." }, "ai-bias": { "label": "helps trace … sources", "note": "It helps separate bias from real causes and surface patterns." } } }, "zh": { "fullName": "结构因果模型", "factExplain": "用变量和因果关系刻画系统的框架。", "humanExplain": "结构因果模型像查寝挨批：晚归、断电、被扣分谁是导火索，谁只是连带遭殃，不能只看总一起发生。\n\n用于分清因果，常见于评估和科研。", "humanExplainDisplay": "结构因果模型像查寝挨批：\n==晚归、断电、被扣分==谁是导火索，\n谁只是连带遭殃，\n不能只看==总一起发生==。\n\n用于分清因果，\n常见于评估和科研。", "relationsNarrative": "Do-calculus\n它提供变量关系与图结构，供 do-calculus 做干预推断。\n\nKnowledge Representation\n它是一种显式表达世界因果关系的知识表示方式。\n\nWorld Model\n若世界模型想表达因果机制，可用它组织状态与作用关系。\n\nAI-bias\n它能帮助分析偏差是由因果机制还是表面相关造成。", "relations": { "do-calculus": { "label": "用…做干预推断", "note": "它给 do-calculus 提供因果结构。" }, "knowledge-representation": { "label": "属于…的一种", "note": "它把因果关系明确编码出来。" }, "world-model": { "label": "可作为…骨架", "note": "世界模型若讲因果，可借它组织变量。" }, "ai-bias": { "label": "帮助分析…来源", "note": "可区分偏差相关性与真实因果。" } } } } }, { "id": "structured-output", "name": "Structured output", "layer": "L4", "era": "2024", "publishedAt": "2026-05-29T16:08:01.211Z", "relations": [ { "to": "function-call" }, { "to": "api" }, { "to": "prompt" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "Structured output", "factExplain": "A way to make AI answer with fixed fields and formats.", "humanExplain": "Structured output is a lunch-order form for AI. It fills the boxes, instead of writing a poem about pizza.\n\nIt makes answers come back in set fields. Apps use it for forms, tool calls, and clean handoffs.", "humanExplainDisplay": "Structured output is a ==lunch-order form== for AI.\nIt ==fills the boxes==,\ninstead of writing a poem about pizza.\n\nIt makes answers come back in set fields.\nApps use it for forms,\ntool calls,\nand clean handoffs.", "relationsNarrative": "Function-calling\nStructured output often lines up the arguments first, then Function-call runs the tool.\n\nAPI\nA stable output shape helps an API or backend catch the result cleanly.\n\nPrompt\nA Prompt often names the fields, types, and return format.\n\nLLM\nStructured output works on the LLM answer, so the model acts neater, not smarter.", "relations": { "function-call": { "label": "sets up …", "note": "It makes fixed arguments first, then Function-call runs the tool." }, "api": { "label": "connects cleanly to …", "note": "A fixed shape helps programs receive the result safely." }, "prompt": { "label": "is guided by …", "note": "The Prompt tells the model which fields to return." }, "llm": { "label": "formats answers from …", "note": "It does not change the brain, only the answer shape." } } }, "zh": { "fullName": "结构化输出", "factExplain": "让模型按预设字段和格式返回结果的输出方式。", "humanExplain": "结构化输出像给 AI 填外卖备注：别自由发挥，按格子写，少放葱也别写成小作文。\n\n它常用于接口返回、表单抽取和工具调用，让程序少猜、多接得上。", "humanExplainDisplay": "结构化输出像==给 AI 填外卖备注==：\n别自由发挥，按格子写，\n少放葱也别写成==小作文==。\n\n它常用于接口返回、\n表单抽取和工具调用，\n让程序少猜、多接得上。", "relationsNarrative": "Function-calling\nStructured output 常先把参数按固定字段排好，再交给 Function-call 去执行。\n\nAPI\n输出结构稳定后，API 和后端程序更容易直接接住结果。\n\nPrompt\nStructured output 往往要靠 Prompt 先说明字段、类型和返回格式。\n\nLLM\n它作用在 LLM 的输出层，重点不是更聪明，而是更规矩。", "relations": { "function-call": { "label": "常配合…执行", "note": "先按格式产出参数，再交给工具处理。" }, "api": { "label": "便于对接…", "note": "结构固定后，程序更容易稳定接收。" }, "prompt": { "label": "通常靠…约束", "note": "提示词会要求模型按指定字段返回。" }, "llm": { "label": "规范…回答格式", "note": "它不改大脑，只约束答案长相。" } } } } }, { "id": "subsumption-architecture", "name": "Subsumption Architecture", "layer": "L3", "era": "1986", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "nouvelle-ai" }, { "to": "robotics" }, { "to": "embodied-ai" }, { "to": "automated-planning" } ], "track": "history", "seo": { "zh": { "title": "包容式架构 是什么?一摞家规便利贴,一文看懂 — AI Rookies", "description": "一种用分层行为直接控制机器人的架构。人话版一分钟讲明白。" }, "en": { "title": "What Is Subsumption Architecture? House Rules on Sticky Notes", "description": "A robot control design with behavior layers driving action directly. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Subsumption Architecture", "factExplain": "A robot control design with behavior layers driving action directly.", "humanExplain": "Subsumption Architecture is like a robot with house rules on sticky notes. “Don’t crash” gets to yell over “go explore.”\n\nYou meet it in mobile robots. They run on stacked reflexes, not one grand plan.", "humanExplainDisplay": "Subsumption Architecture is like a robot\nwith ==house rules== on sticky notes.\n==“Don’t crash”== gets to yell over\n“go explore.”\n\nYou meet it in mobile robots.\nThey run on stacked reflexes,\nnot one grand plan.", "relationsNarrative": "Nouvelle AI\nSubsumption Architecture is a classic control idea in Nouvelle AI.\n\nRobotics\nMobile robots use it to react in behavior layers.\n\nEmbodied AI\nIt lets the robot sense the world and act right away.\n\nPlanning\nIt skips plan-first control and reacts right away.", "relations": { "nouvelle-ai": { "label": "supports …", "note": "It is a classic control idea in Nouvelle AI." }, "robotics": { "label": "used in …", "note": "Early mobile robots used it to react in layers." }, "embodied-ai": { "label": "shows …", "note": "It lets sensing and action stay in a direct loop." }, "automated-planning": { "label": "skips …", "note": "It acts now instead of waiting for a full plan." } } }, "zh": { "fullName": "包容式架构", "factExplain": "一种用分层行为直接控制机器人的架构。", "humanExplain": "包容式架构像一摞家规便利贴：每层规矩都能直接指挥手脚，「别撞墙」永远盖过「去探索」。\n\n用于移动机器人，让机器靠分层反射而非长远空想。", "humanExplainDisplay": "包容式架构像\n==一摞家规便利贴==：\n每层规矩都能直接指挥手脚，\n==「别撞墙」永远盖过「去探索」==。\n\n用于移动机器人，\n让机器靠分层反射，\n而非长远空想。", "relationsNarrative": "Nouvelle AI\n包容式架构是 Nouvelle AI 的代表做法。\n\nRobotics\n它常用于移动机器人，把控制拆成行为层。\n\nEmbodied AI\n它强调身体在环境中直接感知、直接行动。\n\nAutomated Planning\n它绕开先规划再执行，改用即时反应。", "relations": { "nouvelle-ai": { "label": "支撑…", "note": "它是 Nouvelle AI 的代表控制思路。" }, "robotics": { "label": "用于…", "note": "早期移动机器人常靠它分层反应。" }, "embodied-ai": { "label": "体现…", "note": "它强调身体与环境直接闭环。" }, "automated-planning": { "label": "绕开…", "note": "它不依赖完整计划再行动。" } } } } }, { "id": "superintelligence", "name": "Superintelligence", "layer": "L6", "era": "2014", "publishedAt": "2026-05-23T12:00:00Z", "relations": [ { "to": "agi" }, { "to": "singularity" }, { "to": "alignment" }, { "to": "compute-race" } ], "i18n": { "en": { "fullName": "Superintelligence", "factExplain": "An AI system far better than humans at almost every important thinking task.", "humanExplain": "Superintelligence is the classmate with every answer. Five minutes later, they are rewriting recess.\n\nIt means AI could beat humans at almost every hard thinking job. You meet it in serious talks about control and AI safety.", "humanExplainDisplay": "Superintelligence is the classmate\nwith ==every answer==.\nFive minutes later,\nthey are ==rewriting recess==.\n\nIt means AI could beat humans\nat almost every hard thinking job.\nYou meet it in serious talks\nabout control and AI safety.", "relationsNarrative": "AGI\nSuperintelligence is the possible next stage after AGI.\n\nSingularity\nThe Singularity is the big turning point superintelligence might cause.\n\nAlignment\nAlignment helps decide if superintelligence stays under human control.\n\nCompute-race\nCompute-race may speed up the path toward superintelligence.", "relations": { "agi": { "label": "comes after …", "note": "Superintelligence is the possible next stage after AGI." }, "singularity": { "label": "may trigger …", "note": "The Singularity is the turning point superintelligence might create." }, "alignment": { "label": "urgently needs …", "note": "Alignment helps keep superintelligence under human control." }, "compute-race": { "label": "may be sped up by …", "note": "A compute race can push labs faster toward superintelligence." } } }, "zh": { "fullName": "超级智能", "factExplain": "在几乎所有重要认知任务上远超人类的智能系统。", "humanExplain": "超级智能像全班学霸集体开挂，不只抢答，还顺手把题库重写。\n\n它多用于讨论未来风险、对齐和治理，不是今天的产品功能。", "humanExplainDisplay": "超级智能像==全班学霸集体开挂==，\n不只抢答，\n还顺手==把题库重写==。\n\n它多用于讨论未来风险、\n对齐和治理，\n不是今天的产品功能。", "relationsNarrative": "AGI\nSuperintelligence 是 AGI 继续升级后的理论形态。\n\nSingularity\nSingularity 描述 Superintelligence 可能带来的转折点。\n\nAlignment\nAlignment 决定 Superintelligence 是否仍处于可控范围。\n\nCompute-race\nCompute-race 可能加速通向 Superintelligence 的技术进程。", "relations": { "agi": { "label": "是…的下一阶段" }, "singularity": { "label": "可能引发…" }, "alignment": { "label": "亟需…" } } } } }, { "id": "supervised-learning", "name": "Supervised Learning", "layer": "L2", "era": "1960", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "self-supervised-learning" }, { "to": "fine-tuning" }, { "to": "backpropagation" }, { "to": "reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Supervised Learning", "factExplain": "A way to train AI with examples that already have correct answers.", "humanExplain": "Supervised learning is homework with the answer key open. If the AI calls a muffin a cat, the key says, “Nice try, robot.”\n\nIt helps AI sort email. It can also spot faces or predict sales.", "humanExplainDisplay": "Supervised learning is ==homework==\nwith the ==answer key open==.\nIf the AI calls a muffin a cat,\nthe key says, “Nice try, robot.”\n\nIt helps AI sort email.\nIt can also spot faces\nor predict sales.", "relationsNarrative": "SSL\nSupervised Learning uses human labels, while SSL learns from the data itself.\n\nFine-tuning\nFine-tuning often uses labeled data to teach a model a specific task.\n\nBackpropagation\nSupervised Learning often uses Backprop to change the model after errors.\n\nRL\nSupervised Learning learns from correct answers, while RL learns from rewards.", "relations": { "self-supervised-learning": { "label": "uses labels unlike …", "note": "Both train models, but Supervised Learning needs human labels." }, "fine-tuning": { "label": "is often used in …", "note": "Fine-tuning often uses labeled examples to teach a task." }, "backpropagation": { "label": "updates through …", "note": "Supervised Learning often uses Backprop to fix errors." }, "reinforcement-learning": { "label": "learns answers, not … rewards", "note": "Supervised Learning studies correct answers instead of reward scores." } } }, "zh": { "fullName": "监督学习", "factExplain": "用带标签数据训练模型学习输入到输出的映射。", "humanExplain": "跟着驾校教练练车：这道题该打灯、那一步该踩刹车，答案都摆着，错了立刻给你纠回来。\n\n它常用于分类、识别和预测，是很多模型最基础的训练方式。", "humanExplainDisplay": "跟着驾校教练练车：\n这道题该==打灯==、\n那一步该踩刹车，\n答案都摆着，\n错了立刻给你==纠回来==。\n\n它常用于分类、识别和预测，\n是很多模型最基础的训练方式。", "relationsNarrative": "Self-Supervised Learning\n两者都在训练模型，但它明确依赖人工标签。\n\nFine-Tuning\n微调常用带标签数据，把模型拉向具体任务。\n\nBackpropagation\n它通常靠反向传播，根据误差一步步改参数。\n\nReinforcement Learning\n它直接对答案学习，不像强化学习靠奖励试错。", "relations": { "self-supervised-learning": { "label": "对比…看标签", "note": "两者都训练模型，但它明确依赖人工标签。" }, "fine-tuning": { "label": "常被…采用", "note": "微调常用带标签样本教模型学任务。" }, "backpropagation": { "label": "靠…改参数", "note": "它通常用反向传播按误差更新参数。" }, "reinforcement-learning": { "label": "不同于…拿奖励", "note": "它看标准答案，不靠奖励试出来。" } } } } }, { "id": "supply-chain-attack", "name": "Supply Chain Attack", "layer": "L6", "era": "2024", "publishedAt": "2026-05-28T15:58:23.419Z", "relations": [ { "to": "framework" }, { "to": "open-source-model" }, { "to": "data-privacy" }, { "to": "ai-regulation" } ], "track": "today", "i18n": { "en": { "fullName": "Supply Chain Attack", "factExplain": "An attack that sneaks in through trusted software parts or delivery paths.", "humanExplain": "The burglar does not pick your lock. He hides in your new toaster, so you plug him in yourself.\n\nIn software, it hides in dependencies and plugins. It can also ride inside updates, so source and delivery matter.", "humanExplainDisplay": "The burglar does not ==pick your lock==.\nHe hides in your new toaster,\nso you ==plug him in yourself==.\n\nIn software,\nit hides in dependencies and plugins.\nIt can also ride inside updates,\nso source and delivery matter.", "relationsNarrative": "Framework\nSupply chain attacks hide in common framework dependencies and spread into apps.\n\nOpen-source-model\nThey can enter through model downloads, mirrors, or weight files.\n\nData-privacy\nPoisoned dependencies or models can quietly leak user data and keys.\n\nAI-regulation\nThey push source checks, delivery tracking, and clear responsibility.", "relations": { "framework": { "label": "hides in … dependencies", "note": "A common component can carry the attack into many apps." }, "open-source-model": { "label": "slips into … delivery", "note": "Model files and weights can be poisoned before download." }, "data-privacy": { "label": "can threaten …", "note": "Once planted, it can quietly steal data or keys." }, "ai-regulation": { "label": "pushes … audits", "note": "High risk means stronger source checks and clear responsibility." } } }, "zh": { "fullName": "供应链攻击", "factExplain": "通过上游依赖或分发环节间接入侵系统的攻击方式。", "humanExplain": "可怕的不是门口有贼，而是你信任的食材、快递、零件，早在上游就被人悄悄动了手脚。\n\n常藏在依赖、插件和更新链路里，风险重点在来源与分发。", "humanExplainDisplay": "可怕的不是门口有贼，\n而是你信任的食材、快递、零件，\n早在上游就被人\n==悄悄动了手脚==。\n\n常藏在依赖、插件和更新链路里，\n风险重点在来源与分发。", "relationsNarrative": "Framework\n供应链攻击常藏在常用框架、库或其依赖里，借上游组件一路扩散到下游产品。\n\nOpen-source-model\n开源模型的下载、镜像和权重分发链路，也可能成为供应链攻击的入口。\n\nData-privacy\n一旦系统吃进被污染的依赖或模型，用户数据与密钥就可能被悄悄带走。\n\nAI-regulation\n这类风险会推动监管更重视来源审计、分发追踪和责任划分。", "relations": { "framework": { "label": "潜伏在…依赖里", "note": "常借常用组件一路下沉扩散。" }, "open-source-model": { "label": "混入…分发链路", "note": "模型文件与权重来源也可能被动手脚。" }, "data-privacy": { "label": "会威胁…安全", "note": "一旦被植入，数据可能被偷偷带走。" }, "ai-regulation": { "label": "倒逼…加强审计", "note": "风险高时更需要追溯与责任划分。" } } } } }, { "id": "support-vector-machine", "name": "SVM", "layer": "L2", "era": "1995", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "regression" }, { "to": "kernel-method" }, { "to": "statistical-learning-theory" } ], "track": "history", "i18n": { "en": { "fullName": "Support Vector Machine", "factExplain": "A method that learns from labels and separates groups with the widest gap.", "humanExplain": "SVM is like a gym teacher drawing a chalk line between two dodgeball teams. It keeps the widest empty space, and the closest kids matter most.\n\nYou meet it in small classification tasks, often with small data. With a tweak, it can predict numbers too.", "humanExplainDisplay": "SVM is like a gym teacher drawing a ==chalk line==\nbetween two dodgeball teams.\nIt keeps the ==widest empty space==,\nand the closest kids matter most.\n\nYou meet it in small classification tasks,\noften with small data.\nWith a tweak,\nit can predict numbers too.", "relationsNarrative": "Classification\nSVM is best known for classification tasks.\n\nRegression\nA changed SVM can also do regression.\n\nKernel Method\nKernel Method helps SVM split data with curved borders.\n\nSLT\nSLT gives SVM its ideas about wide gaps and new data.", "relations": { "classification": { "label": "is often used for …", "note": "Its classic use is splitting data into two classes." }, "regression": { "label": "can extend to …", "note": "A changed SVM can predict numbers, not just classes." }, "kernel-method": { "label": "uses …", "note": "Kernels let SVM draw curved borders, not just straight ones." }, "statistical-learning-theory": { "label": "is rooted in …", "note": "SLT explains why wide gaps can help on new data." } } }, "zh": { "fullName": "支持向量机", "factExplain": "一种通过最大化分类间隔进行判别的监督学习方法。", "humanExplain": "SVM像小区车位画线，两排车挤得再近，也要找最宽那道缝；贴线最近的车，最有发言权。\n\n常用于小样本分类，也能做回归；数据不大时往往很稳。", "humanExplainDisplay": "SVM像小区车位画线，\n两排车挤得再近，\n也要找最宽那道缝；\n==贴线最近的车==，\n最有==发言权==。\n\n常用于小样本分类，\n也能做回归；\n数据不大时往往很稳。", "relationsNarrative": "Classification\nSVM 最经典的用法，就是做分类任务。\n\nRegression\nSVM 经过改造后，也可以用于回归预测。\n\nKernel Method\n核方法让它能处理原本分不开的非线性数据。\n\nStatistical Learning Theory\nSVM 的核心思想，与间隔和泛化理论紧密相关。", "relations": { "classification": { "label": "常用于…任务", "note": "它最经典的用途就是二分类。" }, "regression": { "label": "也可扩展到…", "note": "改造后可用于连续值预测。" }, "kernel-method": { "label": "属于…代表", "note": "核技巧让它能处理非线性边界。" }, "statistical-learning-theory": { "label": "扎根于…", "note": "它和间隔、泛化理论关系很深。" } } } } }, { "id": "symbolic-ai", "name": "Symbolic AI", "layer": "L1", "era": "1960", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "knowledge-representation" }, { "to": "logic-programming" }, { "to": "automated-planning" }, { "to": "connectionism" } ], "track": "history", "i18n": { "en": { "fullName": "Symbolic AI", "factExplain": "An AI style using symbols and rules to store knowledge and reason.", "humanExplain": "Symbolic AI is a board-game referee with a tiny whistle. No vibes, no guessing, just the rulebook.\n\nIt works well when rules are clear, like planning steps. It can freeze when common sense gets fuzzy.", "humanExplainDisplay": "Symbolic AI is a ==board-game referee==\nwith a tiny whistle.\n==No vibes, no guessing==,\njust the rulebook.\n\nIt works well when rules are clear,\nlike planning steps.\nIt can freeze when common sense gets fuzzy.", "relationsNarrative": "KR\nSymbolic AI uses KR to write the world as knowledge it can reason with.\n\nLogic\nSymbolic AI uses Logic to turn rules into programs.\n\nPlanning\nPlanning is one of Symbolic AI’s classic jobs.\n\nConnectionism\nSymbolic AI and Connectionism split the rule camp from the neural network camp.", "relations": { "knowledge-representation": { "label": "represents knowledge with …", "note": "KR is the base for symbolic reasoning." }, "logic-programming": { "label": "writes rules with …", "note": "Logic turns rules into runnable reasoning." }, "automated-planning": { "label": "supports …", "note": "Planning often uses symbolic states and action rules." }, "connectionism": { "label": "contrasts with …", "note": "One uses clear rules. The other uses neural connections." } } }, "zh": { "fullName": "符号主义人工智能", "factExplain": "用符号和规则表示知识并推理的 AI 范式。", "humanExplain": "符号主义把AI练成武林弟子：不靠灵感靠规矩，招式口诀写明白，出手全按门规来。\n\n它适合清晰规则的推理和规划；遇到模糊常识易卡。", "humanExplainDisplay": "符号主义把AI练成\n==武林弟子==：\n不靠灵感靠规矩，\n招式口诀写明白，出手全按==门规==来。\n\n它适合清晰规则的推理和规划；\n遇到模糊常识易卡。", "relationsNarrative": "Knowledge Representation\n符号主义常靠它把世界写成可推理的知识。\n\nLogic Programming\n它用逻辑规则把符号推理写成程序。\n\nAutomated Planning\n规划是符号主义最典型的任务之一。\n\nConnectionism\n它们代表规则派与神经网络派的分野。", "relations": { "knowledge-representation": { "label": "用…表示知识", "note": "知识表示是符号推理的地基。" }, "logic-programming": { "label": "用…写规则", "note": "逻辑程序把规则变成可执行推理。" }, "automated-planning": { "label": "支撑…", "note": "规划常用符号状态和动作规则。" }, "connectionism": { "label": "对照…", "note": "一个靠明规则，一个靠神经连接。" } } } } }, { "id": "syntactic-parsing", "name": "Syntax Parse", "layer": "L4", "era": "1950s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "natural-language-understanding" }, { "to": "information-extraction" }, { "to": "machine-translation" } ], "track": "history", "i18n": { "en": { "fullName": "Syntactic Parsing", "factExplain": "A language method that finds the grammar structure of a sentence.", "humanExplain": "Syntactic parsing is a seating chart for a sentence. It stops the verb from dancing with the wrong noun.\n\nIt finds the sentence skeleton first. Search, translation, and fact-pulling tools use that skeleton.", "humanExplainDisplay": "Syntactic parsing is a ==seating chart== for a sentence.\nIt stops the verb from ==dancing with the wrong noun==.\n\nIt finds the sentence skeleton first.\nSearch, translation, and fact-pulling tools use that skeleton.", "relationsNarrative": "NLP\nSyntactic Parsing is a classic basic task in NLP.\n\nNLU\nIt pulls out the sentence skeleton and helps NLU read the meaning.\n\nIE\nIt helps IE avoid mixing up entities and their relations.\n\nMT\nIt helps MT keep the grammar structure of the original sentence.", "relations": { "natural-language-processing": { "label": "is a basic task in …", "note": "It was one of the classic early tasks in NLP." }, "natural-language-understanding": { "label": "helps … understand meaning", "note": "Grammar structure helps a system understand sentence meaning." }, "information-extraction": { "label": "helps … find relations", "note": "Clear sentence parts make entity relations easier to extract." }, "machine-translation": { "label": "helps … keep structure", "note": "Translation often needs the old sentence structure to stay clear." } } }, "zh": { "fullName": "Syntactic Parsing（句法分析）", "factExplain": "识别句子语法结构的语言处理方法。", "humanExplain": "句法分析像相亲局拆关系：谁主谁宾、谁修饰谁，别把七大姑当对象。\n\n用于搜索、翻译和信息抽取，先理清句子骨架。", "humanExplainDisplay": "句法分析像相亲局\n==拆关系==：\n谁主谁宾、谁修饰谁，\n别把七大姑当对象。\n\n用于搜索、翻译，\n和信息抽取，\n先理清句子骨架。", "relationsNarrative": "NLP\nSyntactic Parsing 是 NLP 里的经典基础任务。\n\nNLU\n它把句子骨架拆出来，帮助理解句意。\n\nInformation Extraction\n它让实体和关系抽取少看错搭配。\n\nMachine Translation\n翻译常借它保留原句的语法结构。", "relations": { "natural-language-processing": { "label": "属于…基础任务", "note": "它是早期 NLP 的核心任务之一。" }, "natural-language-understanding": { "label": "支撑…理解句意", "note": "语法结构能帮助理解句意。" }, "information-extraction": { "label": "帮…定位关系", "note": "先拆清句子，抽实体关系更稳。" }, "machine-translation": { "label": "服务…保留结构", "note": "翻译时常要保留原句结构。" } } } } }, { "id": "synthetic-data", "name": "Synthetic Data", "layer": "L2", "era": "1990s", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "data-augmentation" }, { "to": "generative-model" }, { "to": "data-privacy" }, { "to": "ai-bias" } ], "track": "today", "i18n": { "en": { "fullName": "Synthetic Data", "factExplain": "Data made by a program or model for training or testing AI.", "humanExplain": "Synthetic data is like a toy city for self-driving cars. The cars can practice without denting your mailbox.\n\nPeople use it to train AI, test AI, and protect private data. But if the toy city is weird, the AI learns weird habits.", "humanExplainDisplay": "Synthetic data is like a ==toy city==\nfor self-driving cars.\nThe cars can practice\nwithout ==denting your mailbox==.\n\nPeople use it to train AI,\ntest AI,\nand protect private data.\nBut if the toy city is weird,\nthe AI learns weird habits.", "relationsNarrative": "Data Augmentation\nSynthetic data often adds more examples when real data is scarce.\n\nGenerative Model\nA generative model can make synthetic data in large batches.\n\nData-privacy\nSynthetic data can reduce exposure of sensitive real data.\n\nAI-bias\nIf synthetic data is skewed, it can feed bias into the model.", "relations": { "data-augmentation": { "label": "often used for …", "note": "Synthetic data fills gaps when real examples are rare." }, "generative-model": { "label": "often made by …", "note": "Generative models can create many realistic fake examples." }, "data-privacy": { "label": "can lower … risk", "note": "It avoids showing the original sensitive data directly." }, "ai-bias": { "label": "can amplify …", "note": "Bad fake data can teach the model the same bias, or worse." } } }, "zh": { "fullName": "合成数据", "factExplain": "由程序或模型生成、用于训练或测试的数据。", "humanExplain": "合成数据像驾校模拟路考：真马路不够练，先摆雪糕筒造场景。\n\n用于训练、测试和隐私保护，造偏了模型也跑偏。", "humanExplainDisplay": "合成数据像驾校模拟路考：\n真马路==不够练==，\n先摆雪糕筒，\n==造场景==。\n\n用于训练、测试，\n和隐私保护，\n造偏了模型也跑偏。", "relationsNarrative": "Data Augmentation\n合成数据常被用来扩充稀缺样本。\n\nGenerative Model\n生成模型能批量造出合成数据。\n\nData Privacy\n合成数据可减少暴露敏感原始数据。\n\nAI Bias\n合成数据仿歪时，会把偏见喂进模型。", "relations": { "data-augmentation": { "label": "常用于…", "note": "合成数据能补足真实样本空缺。" }, "generative-model": { "label": "常由…生成", "note": "生成模型负责造出逼真样本。" }, "data-privacy": { "label": "可降低…风险", "note": "不用直接暴露原始敏感数据。" }, "ai-bias": { "label": "可能放大…", "note": "生成过程可能继承甚至放大偏见。" } } } } }, { "id": "system-prompt", "name": "System prompt", "layer": "L5", "sublayer": "architecture", "era": "2020s", "publishedAt": "2026-06-02T04:00:00.000Z", "relations": [ { "to": "prompt" }, { "to": "prompt-injection" }, { "to": "agent" }, { "to": "alignment" } ], "track": "today", "i18n": { "en": { "fullName": "System prompt", "factExplain": "A hidden instruction that sets the AI’s role and rules before the chat starts.", "humanExplain": "It is like a note taped behind the stage. The actors cannot see it. But it tells them how to act.\n\nIt keeps the AI’s tone and limits steady. You often meet it in assistants, customer support, and Agents.", "humanExplainDisplay": "It is like a ==note taped behind the stage==.\nThe actors cannot see it.\nBut it tells them ==how to act==.\n\nIt keeps the AI’s tone and limits steady.\nYou often meet it in assistants,\ncustomer support,\nand Agents.", "relationsNarrative": "Prompt\nThe System prompt sets the model’s role and rules before the user Prompt.\n\nPrompt injection\nPrompt injection often tries to trick the model into breaking the System prompt.\n\nAgent\nThe System prompt often sets an Agent’s behavior limits and style.\n\nAlignment\nSome Alignment rules are put into practice through the System prompt.", "relations": { "prompt": { "label": "sets rules before …", "note": "It comes before the user Prompt and acts like the base playbook." }, "prompt-injection": { "label": "can be bypassed by …", "note": "Bad inputs often try to make the model ignore these rules." }, "agent": { "label": "constrains … behavior", "note": "An Agent’s goals, tone, and limits are often set here." }, "alignment": { "label": "puts … into practice", "note": "Many safety and behavior rules are carried through the system prompt." } } }, "zh": { "fullName": "系统提示词", "factExplain": "在对话前为模型设定角色与规则的隐藏指令。", "humanExplain": "别看它不露脸，其实更像后台导演：台词能说到哪、角色怎么演，先把 AI 按进剧本里。\n\n它用来统一语气和边界，常见于助手、客服与 Agent 场景。", "humanExplainDisplay": "别看它不露脸，\n其实更像==后台导演==：\n台词能说到哪、\n角色怎么演，\n先把 AI 按进==剧本里==。\n\n它用来统一语气和边界，\n常见于助手、\n客服与 Agent 场景。", "relationsNarrative": "Prompt\nSystem prompt 在用户 Prompt 之前，先给模型设定角色和规则。\n\nPrompt injection\nPrompt injection 常试图诱导模型违背 System prompt。\n\nAgent\nAgent 的行为边界和执行风格常由 System prompt 约束。\n\nAlignment\nAlignment 的部分安全要求，常通过 System prompt 落地。", "relations": { "prompt": { "label": "先于…定规矩", "note": "它比用户 Prompt 更像底层操作手册。" }, "prompt-injection": { "label": "会被…绕过", "note": "恶意输入常试图让模型忽略系统规则。" }, "agent": { "label": "约束…行为", "note": "Agent 的目标、语气和边界常靠它设定。" }, "alignment": { "label": "承接…要求", "note": "很多安全与行为约束会落到这里实现。" } } } } }, { "id": "t5", "name": "T5", "layer": "L3", "era": "2019", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "seq2seq" }, { "to": "pretraining" }, { "to": "fine-tuning" } ], "track": "history", "i18n": { "en": { "fullName": "Text-to-Text Transfer Transformer", "factExplain": "A Transformer model handling many language tasks as text in and text out.", "humanExplain": "T5 is like a drive-thru window for words. You hand over text. It hands back text.\n\nIt uses one setup for many language jobs. It can summarize, translate, and answer questions.", "humanExplainDisplay": "T5 is like a ==drive-thru window== for words.\nYou hand over ==text==.\nIt hands back text.\n\nIt uses one setup for many language jobs.\nIt can summarize, translate, and answer questions.", "relationsNarrative": "Transformer\nT5 is an encoder-decoder Transformer model.\n\nSeq2Seq\nT5 turns many language jobs into Seq2Seq jobs.\n\nPretraining\nT5 first learns language skills from large text pretraining.\n\nFine-tuning\nFine-tuning adapts T5 for summary, translation, and Q&A jobs.", "relations": { "transformer": { "label": "built on …", "note": "Transformer gives T5 its main model shape." }, "seq2seq": { "label": "uses … form", "note": "T5 treats input and output as text sequences." }, "pretraining": { "label": "starts with …", "note": "Pretraining gives T5 broad language basics." }, "fine-tuning": { "label": "adapts with …", "note": "Fine-tuning fits T5 to a specific job." } } }, "zh": { "fullName": "Text-to-Text Transfer Transformer，文本到文本迁移 Transformer", "factExplain": "把各类语言任务统一成文本到文本的 Transformer 模型。", "humanExplain": "T5像食堂只收饭票：打饭、买面、加菜，统统先换成饭票办。\n\n翻译、摘要、问答，全走\"文字进文字出\"一条道。", "humanExplainDisplay": "T5像食堂\n==只收饭票==：\n打饭、买面、加菜，\n统统先==换成饭票==办。\n\n翻译、摘要、问答，\n全走\"文字进文字出\"一条道。", "relationsNarrative": "Transformer\nT5 是编码器-解码器式的 Transformer 模型。\n\nSeq2Seq\nT5 把各种任务都改写成序列到序列问题。\n\nPretraining\nT5 先靠大规模文本预训练获得语言能力。\n\nFine-tuning\n微调让 T5 适配摘要、翻译、问答等任务。", "relations": { "transformer": { "label": "基于…架构", "note": "Transformer 给它提供主体架构。" }, "seq2seq": { "label": "采用…形式", "note": "输入和输出都被当成文本序列。" }, "pretraining": { "label": "通过…打底", "note": "预训练提供通用语言底子。" }, "fine-tuning": { "label": "用…适配任务", "note": "微调把它适配到具体任务。" } } } } }, { "id": "tabular-foundation-model", "name": "Tabular Foundation Model", "layer": "L3", "era": "2022", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "foundation-model" }, { "to": "feature-engineering" }, { "to": "xgboost" }, { "to": "transformer" } ], "track": "today", "seo": { "en": { "title": "What Are Tabular Foundation Models? Pretrained AI for Table Data", "description": "A foundation model that has seen thousands of tables: show it a new one and it starts predicting, with little feature engineering. Explained in plain English." }, "zh": { "title": "表格基础模型是什么?看过千百张报表的老会计,一文看懂 — AI Rookies", "description": "新表一摊,先认列再开算:面向表格数据预训练、可迁移复用,少造特征也能起步。人话版一分钟讲明白。" } }, "i18n": { "en": { "fullName": "Tabular Foundation Model", "factExplain": "A pretrained foundation model built for table data and reused on new tables.", "humanExplain": "Think of a spreadsheet whiz who has read every budget sheet in town. Give it a new table, and it spots the columns before grabbing a calculator.\n\nIt helps predict numbers. It helps check risk. It helps teams study daily work. It can start with fewer hand-made features.", "humanExplainDisplay": "Think of a ==spreadsheet whiz==\nwho has read every budget sheet in town.\nGive it a new table,\nand it ==spots the columns==\nbefore grabbing a calculator.\n\nIt helps predict numbers.\nIt helps check risk.\nIt helps teams study daily work.\nIt can start with fewer hand-made features.", "relationsNarrative": "Foundation-model\nA Tabular Foundation Model uses large pretraining first, then moves to new tables.\n\nFeature-engineering\nIt tries to learn rows and columns, so people make fewer features by hand.\n\nXGBoost\nXGBoost is the old champ for table prediction and a key benchmark.\n\nTransformer\nMany Tabular Foundation Models use Transformers to read rows and columns.", "relations": { "foundation-model": { "label": "is a kind of …", "note": "It brings the foundation model idea to table data." }, "feature-engineering": { "label": "needs less …", "note": "It learns column links, so people make fewer features by hand." }, "xgboost": { "label": "challenges …", "note": "XGBoost has long been the strong player for table prediction." }, "transformer": { "label": "often uses …", "note": "Many tabular foundation models use Transformers." } } }, "zh": { "fullName": "表格基础模型", "factExplain": "面向表格数据预训练、可迁移的基础模型。", "humanExplain": "表格基础模型像看过千百张报表的老会计：新表一摊，先认列再开算。\n\n用于预测、风控和运营分析，少造特征也能起步。", "humanExplainDisplay": "表格基础模型像\n看过==千百张报表==的老会计：\n新表一摊，\n==先认列再开算==。\n\n用于预测、风控和运营分析，\n少造特征也能起步。", "relationsNarrative": "Foundation-model\n它把“先大规模预训练、再迁移”的思路用于表格。\n\nFeature-engineering\n它希望自动理解列与行，减少手工造特征。\n\nXGBoost\nXGBoost 是表格预测的传统强手，也是主要对照。\n\nTransformer\n许多表格基础模型用 Transformer 处理列、行关系。", "relations": { "foundation-model": { "label": "属于…的一类", "note": "把基础模型思路搬到表格数据上。" }, "feature-engineering": { "label": "减少…依赖", "note": "自动理解列关系，少靠手工特征。" }, "xgboost": { "label": "挑战…的地盘", "note": "表格预测长期由 XGBoost 占优。" }, "transformer": { "label": "常用…建模", "note": "许多表格基础模型采用 Transformer。" } } } } }, { "id": "td-gammon", "name": "TD-Gammon", "layer": "L4", "era": "1992", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "temporal-difference-learning" }, { "to": "reinforcement-learning" }, { "to": "neural-network" }, { "to": "alphazero" } ], "track": "history", "i18n": { "en": { "fullName": "Temporal Difference Gammon", "factExplain": "A backgammon AI trained by playing itself with TD learning.", "humanExplain": "TD-Gammon is like a kid playing both sides at lunch. It loses, groans, swaps seats, and gets smarter.\n\nIt used a neural network to judge backgammon boards. It showed self-play could train strong game AI.", "humanExplainDisplay": "TD-Gammon is like a kid\n==playing both sides== at lunch.\nIt loses, groans,\n==swaps seats==,\nand gets smarter.\n\nIt used a neural network\nto judge backgammon boards.\nIt showed self-play could train\nstrong game AI.", "relationsNarrative": "TD Learning\nTD-Gammon used TD Learning to update its board scores.\n\nRL\nTD-Gammon was an early famous win for RL.\n\nNeural-network\nTD-Gammon used a Neural-network to estimate its chance of winning.\n\nAlphaZero\nAlphaZero followed the self-play path that TD-Gammon helped show.", "relations": { "temporal-difference-learning": { "label": "updates play with …", "note": "TD Learning told it if a board position got better or worse." }, "reinforcement-learning": { "label": "was an early win for …", "note": "It improved its moves from wins and losses." }, "neural-network": { "label": "judges boards with …", "note": "Its neural network turned a board into a win estimate." }, "alphazero": { "label": "helped inspire …", "note": "Both got stronger by playing against themselves." } } }, "zh": { "fullName": "时序差分双陆棋程序", "factExplain": "用时序差分自我对弈训练的双陆棋神经网络。", "humanExplain": "TD-Gammon 像游戏开影分身：自己陪自己刷局，每走一步就打分，不等终局才复盘。\n\n证明自我对弈能练策略，影响后来的游戏 AI。", "humanExplainDisplay": "TD-Gammon 像游戏\n开==影分身==：\n自己陪自己刷局，\n==每走一步就打分==，\n不等终局才复盘。\n\n证明自我对弈能练策略，\n影响后来的游戏 AI。", "relationsNarrative": "TD Learning\nTD-Gammon 用 TD 误差更新局面价值评估。\n\nReinforcement Learning\nTD-Gammon 是强化学习早期出圈的成功案例。\n\nNeural Network\nTD-Gammon 用神经网络估计棋盘局面的胜率。\n\nAlphaZero\nAlphaZero 延续了自我对弈强化学习的路线。", "relations": { "temporal-difference-learning": { "label": "用…更新棋感", "note": "它用 TD 误差评估局面好坏。" }, "reinforcement-learning": { "label": "是…早期范例", "note": "它靠输赢反馈改进下棋策略。" }, "neural-network": { "label": "用…评估局面", "note": "神经网络把棋盘映成胜率。" }, "alphazero": { "label": "启发…自我对弈", "note": "两者都靠自我对弈变强。" } } } } }, { "id": "teacher-forcing", "name": "Teacher Forcing", "layer": "L2", "era": "1989", "publishedAt": "2026-07-11T04:00:00.000Z", "relations": [ { "to": "seq2seq" }, { "to": "autoregressive-model" }, { "to": "recurrent-neural-network" }, { "to": "cross-entropy-loss" } ], "track": "history", "seo": { "zh": { "title": "教师强制 是什么?老师带读课文,一文看懂 — AI Rookies", "description": "训练序列模型时喂真实上一步输出。人话版一分钟讲明白。" }, "en": { "title": "What Is Teacher Forcing? Karaoke With Bossy Lyrics", "description": "A training trick that feeds the real previous answer to a sequence model. Explained simply, with a concrete analogy and a concept map." } }, "i18n": { "en": { "fullName": "Teacher Forcing", "factExplain": "A training trick that feeds the real previous answer to a sequence model.", "humanExplain": "Teacher Forcing is karaoke with bossy lyrics. You sing 'banana' by mistake, but the screen still shows the real next line.\n\nDuring training, the model gets the correct last step, not its own guess. This makes training faster and steadier in translation, chat, and speech AI.", "humanExplainDisplay": "Teacher Forcing is ==karaoke== with ==bossy lyrics==.\nYou sing 'banana' by mistake,\nbut the screen still shows the real next line.\n\nDuring training,\nthe model gets the correct last step,\nnot its own guess.\nThis makes training faster and steadier\nin translation, chat, and speech AI.", "relationsNarrative": "Seq2Seq\nTeacher Forcing often trains the decoder and keeps early guesses from going wild.\n\nAutoregressive Model\nIt teaches the model with real past steps before asking for the next one.\n\nRNN\nIt was often used for early RNN sequence prediction.\n\nCross-Entropy Loss\nThe real next token is usually used to compute Cross-Entropy Loss.", "relations": { "seq2seq": { "label": "steadies … training", "note": "Seq2Seq uses it to keep the decoder from drifting early." }, "autoregressive-model": { "label": "trains … with real history", "note": "It gives the model the real past step while it learns." }, "recurrent-neural-network": { "label": "helps … learn sequences", "note": "Early RNN sequence training often used this trick." }, "cross-entropy-loss": { "label": "pairs with …", "note": "The real next token is often used to compute cross-entropy." } } }, "zh": { "fullName": "教师强制", "factExplain": "训练序列模型时喂真实上一步输出。", "humanExplain": "教师强制像老师带读课文：你刚念跑偏，下一句仍把原文怼到眼前。\n\n让序列训练更稳更快，常用于翻译、对话和语音。", "humanExplainDisplay": "教师强制像老师带读课文：\n你刚==念跑偏==，\n下一句仍把==原文怼到眼前==。\n\n让序列训练更稳更快，\n常用于翻译、对话和语音。", "relationsNarrative": "Seq2Seq\n教师强制常用于训练解码器，减少早期乱跑。\n\nAutoregressive Model\n它用真实上文教逐步生成模型预测下一步。\n\nRNN\n它最早常用于 RNN 的序列预测训练。\n\nCross-Entropy Loss\n真实下一步通常直接拿来计算交叉熵。", "relations": { "seq2seq": { "label": "稳定训练…", "note": "Seq2Seq 常用它稳住解码器。" }, "autoregressive-model": { "label": "喂真值训练…", "note": "它用真值上文训练逐步生成。" }, "recurrent-neural-network": { "label": "帮助…学序列", "note": "早期常用于 RNN 序列预测。" }, "cross-entropy-loss": { "label": "配合…优化", "note": "真值 token 常用来算交叉熵。" } } } } }, { "id": "temperature", "name": "Temperature", "layer": "L2", "era": "2010s", "publishedAt": "2026-05-23T09:15:00Z", "relations": [ { "to": "prompt" }, { "to": "inference" }, { "to": "llm" } ], "i18n": { "en": { "fullName": "Temperature (Sampling)", "factExplain": "A setting that controls how random or creative an AI answer is.", "humanExplain": "Temperature is the AI’s glitter dial. Low means a neat worksheet. High means glitter on the dog.\n\nYou meet it in AI settings. Set it low for facts, or higher for ideas. Too high gets wild.", "humanExplainDisplay": "Temperature is the AI’s ==glitter dial==.\nLow means a ==neat worksheet==.\nHigh means glitter on the dog.\n\nYou meet it in AI settings.\nSet it low for facts,\nor higher for ideas.\nToo high gets wild.", "relationsNarrative": "Prompt\nThe same Prompt can give different answers when Temperature changes.\n\nInference\nTemperature controls randomness during Inference.\n\nLLM\nHigher Temperature makes LLM output more varied and less steady.", "relations": { "prompt": { "label": "works with …", "note": "The same Prompt can give different answers at different Temperature settings." }, "inference": { "label": "takes effect during …", "note": "Temperature controls sampling randomness during Inference." }, "llm": { "label": "tunes … output", "note": "Higher Temperature makes LLM output more varied and less steady." } } }, "zh": { "fullName": "温度（采样）", "factExplain": "控制模型输出随机性和创造性的采样参数。", "humanExplain": "温度像 AI 的酒量旋钮，调低像写公文，调高像凌晨三点灵感上头。\n\n做事实问答要低一点，写创意文案可以高一点，但别让它放飞到没边。", "humanExplainDisplay": "温度像 AI 的\n==酒量旋钮==。\n调低像写公文，\n调高像凌晨三点灵感上头。\n\n事实问答适合低一点，\n创意文案可以高一点。\n再高，就容易开始胡言乱语。", "relationsNarrative": "Prompt\n同一个 Prompt 会因 Temperature 不同而生成不同结果。\n\nInference\nTemperature 在 Inference 阶段控制采样随机性。\n\nLLM\nTemperature 越高，LLM 输出越多样也越不稳定。", "relations": { "prompt": { "label": "配合…使用" }, "inference": { "label": "在…时生效" }, "llm": { "label": "调节…的输出" } } } } }, { "id": "temporal-difference-learning", "name": "TD Learning", "layer": "L2", "era": "1988", "publishedAt": "2026-06-05T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "q-learning" }, { "to": "actor-critic" }, { "to": "policy-gradient" } ], "track": "history", "i18n": { "en": { "fullName": "Temporal-Difference Learning", "factExplain": "A reinforcement learning method that updates value guesses step by step.", "humanExplain": "It is like fixing your pancake recipe after each pancake. Too pale? Turn up the heat before breakfast is over.\n\nTD Learning updates an agent’s value guesses while it acts. It helps the agent learn as it goes.", "humanExplainDisplay": "It is like fixing your pancake recipe\n==after each pancake==.\nToo pale?\nTurn up the heat\n==before breakfast is over==.\n\nTD Learning updates an agent’s value guesses\nwhile it acts.\nIt helps the agent learn as it goes.", "relationsNarrative": "RL\nTD Learning is a classic RL method for updating value guesses step by step.\n\nQ-Learning\nQ-Learning uses TD Learning to update Q values before the whole run ends.\n\nActor-Critic\nActor-Critic often uses TD Learning to keep its critic improving.\n\nPolicy Gradient\nTD Learning learns value guesses. Policy Gradient changes the policy more directly.", "relations": { "reinforcement-learning": { "label": "is a core method in …", "note": "TD Learning is a classic way RL updates value guesses." }, "q-learning": { "label": "powers … updates", "note": "Q-Learning uses TD updates before the whole run ends." }, "actor-critic": { "label": "often trains …", "note": "Actor-Critic often uses TD Learning for its critic." }, "policy-gradient": { "label": "complements …", "note": "TD Learning learns values. Policy Gradient changes the policy directly." } } }, "zh": { "fullName": "时序差分学习（Temporal-Difference Learning）", "factExplain": "一种用当前估计差值逐步更新价值的强化学习方法。", "humanExplain": "它像月考后估分：不用等期末总成绩，先看这题比预想高还是低，下一次立刻改策略。\n\n常用来更新价值判断，让智能体边走边学、越估越准。", "humanExplainDisplay": "它像月考后估分：\n不用等==期末总成绩==，\n先看这题比预想高还是低，\n下一次立刻==改策略==。\n\n常用来更新价值判断，\n让智能体边走边学、\n越估越准。", "relationsNarrative": "Reinforcement-learning\n它是强化学习中的经典学习方法，用奖励差值逐步更新判断。\n\nQ-learning\nQ-learning 用它更新 Q 值，不必等整局结束再学习。\n\nActor-critic\nActor-Critic 里负责评价的 critic，常靠它持续修正。\n\nPolicy-gradient\n它重在学价值估计，和直接优化策略的方法形成互补。", "relations": { "reinforcement-learning": { "label": "属于…核心方法", "note": "它是强化学习里最经典的更新思路之一。" }, "q-learning": { "label": "支撑…更新", "note": "Q-learning 就建立在时序差分更新上。" }, "actor-critic": { "label": "常被…采用", "note": "Actor-Critic 常用它来训练价值部分。" }, "policy-gradient": { "label": "与…互补", "note": "它偏价值估计，后者偏直接改策略。" } } } } }, { "id": "tensorflow", "name": "TensorFlow", "layer": "L5", "sublayer": "product", "era": "2015", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "framework" }, { "to": "pytorch" }, { "to": "automatic-differentiation" }, { "to": "gpu" } ], "track": "history", "i18n": { "en": { "fullName": "TensorFlow", "factExplain": "An open-source framework for building and training machine learning models.", "humanExplain": "TensorFlow is a power-tool kit for building AI. You build the brain, train it, and ship it.\n\nPeople use it to train models and put them into apps. It also runs on many kinds of devices.", "humanExplainDisplay": "TensorFlow is a ==power-tool kit== for building AI.\nYou build the brain,\ntrain it,\nand ==ship it==.\n\nPeople use it to train models\nand put them into apps.\nIt also runs on many kinds of devices.", "relationsNarrative": "Framework\nTensorFlow is a well-known example of a machine learning framework.\n\nPyTorch\nTensorFlow is often compared with PyTorch for ease of use and tools.\n\nAutodiff\nTensorFlow uses Autodiff to train neural networks more easily.\n\nGPU\nTensorFlow often uses GPUs to speed up training and math.", "relations": { "framework": { "label": "is a classic …", "note": "TensorFlow is one of the classic machine learning frameworks." }, "pytorch": { "label": "is compared with …", "note": "TensorFlow and PyTorch are both major deep learning frameworks." }, "automatic-differentiation": { "label": "has built-in …", "note": "Autodiff helps TensorFlow train neural networks." }, "gpu": { "label": "speeds up with …", "note": "GPUs help TensorFlow train large models faster." } } }, "zh": { "fullName": "谷歌开源的机器学习框架", "factExplain": "一个用于构建和训练机器学习模型的开源框架。", "humanExplain": "做 AI 搭系统，它像一套全家桶装修工具：砌网络、拉训练、接部署，基本都能一条龙。\n\n常用于训练和部署模型，适合研究、工业落地，以及跨设备上线。", "humanExplainDisplay": "做 AI 搭系统，\n它像一套\n==全家桶装修工具==：\n砌网络、拉训练、接部署，\n基本都能\n==一条龙==。\n\n常用于训练和部署模型，\n适合研究、工业落地，\n以及跨设备上线。", "relationsNarrative": "Framework\nTensorFlow 是机器学习框架的代表性实现。\n\nPyTorch\n它常与 PyTorch 并列，被拿来比较易用性与生态。\n\nAutomatic Differentiation\n它内置自动求导，方便训练神经网络。\n\nGPU\n它通常借助 GPU 加速模型训练与计算。", "relations": { "framework": { "label": "属于…代表", "note": "它是经典机器学习框架之一。" }, "pytorch": { "label": "常与…对比", "note": "两者长期是主流深度学习框架。" }, "automatic-differentiation": { "label": "内置…能力", "note": "自动求导是训练神经网络的基础。" }, "gpu": { "label": "依赖…加速", "note": "训练大模型时常需 GPU 提速。" } } } } }, { "id": "test-time-compute", "name": "TTC", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-13T04:00:00.000Z", "relations": [ { "to": "reasoning-effort" }, { "to": "inference" }, { "to": "reasoning-model" }, { "to": "compute-race" } ], "track": "today", "i18n": { "en": { "fullName": "Test-time compute", "factExplain": "Extra computing power an AI uses while making an answer.", "humanExplain": "TTC is like grabbing extra scratch paper on a math test. You did not get smarter, but you stopped guessing and worked it out.\n\nYou see it in hard reasoning tasks. You also see it in code and math. More compute can make answers steadier. It also takes more time and money.", "humanExplainDisplay": "TTC is like grabbing ==extra scratch paper==\non a math test.\nYou did not get smarter,\nbut you stopped guessing\nand ==worked it out==.\n\nYou see it in hard reasoning tasks.\nYou also see it in code and math.\nMore compute can make answers steadier.\nIt also takes more time and money.", "relationsNarrative": "Reasoning effort\nReasoning effort often controls how much extra thinking the model uses.\n\nInference\nTTC happens during inference, not during training.\n\nReasoning-model\nA reasoning model can spend more compute at answer time for a steadier answer.\n\nCompute-race\nIf every model thinks longer, the compute-race gets hotter.", "relations": { "reasoning-effort": { "label": "is set by …", "note": "Reasoning effort is the knob for more or less TTC." }, "inference": { "label": "happens during …", "note": "TTC adds compute while the model is answering." }, "reasoning-model": { "label": "boosts …", "note": "Reasoning models often use more TTC for steadier answers." }, "compute-race": { "label": "adds pressure to …", "note": "More answer-time compute means more demand for chips and power." } } }, "zh": { "fullName": "Test-time compute（测试时计算量）", "factExplain": "模型在生成答案时额外投入的计算资源。", "humanExplain": "这就像下棋时肯不肯多停几手：不是棋手突然变强了，而是愿意==多算几步后手==，少来点==拍脑袋落子==。\n\n常见于推理、代码和数学题；算得更久，通常更稳，但也更慢更贵。", "humanExplainDisplay": "这就像下棋时肯不肯多停几手：\n不是棋手突然变强了，\n而是愿意==多算几步后手==，\n少来点==拍脑袋落子==。\n\n常见于推理、\n代码和数学题；\n算得更久，\n通常更稳，但也更慢更贵。", "relationsNarrative": "Reasoning effort\n它常通过“思考更多”这类档位被直接调节。\n\nInference\n它发生在出答案阶段，不是训练阶段加算力。\n\nReasoning-model\n推理模型常靠临场多算几步，换更稳的答案。\n\nCompute-race\n如果大家都让模型临场多算，算力压力也会更大。", "relations": { "reasoning-effort": { "label": "常由…调节", "note": "常用努力档位控制它的多少。" }, "inference": { "label": "发生在…阶段", "note": "它增加的是出答案时的计算。" }, "reasoning-model": { "label": "支撑…表现", "note": "推理模型常靠它换更强表现。" }, "compute-race": { "label": "加剧…压力", "note": "更多在线计算会推高算力需求。" } } } } }, { "id": "text-classification", "name": "Text Cls", "layer": "L4", "era": "1960s", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "classification" }, { "to": "natural-language-processing" }, { "to": "supervised-learning" }, { "to": "bert" } ], "track": "history", "i18n": { "en": { "fullName": "Text Classification", "factExplain": "An NLP task for putting text into preset labels automatically.", "humanExplain": "Text classification is a mailroom clerk with a label maker. It stamps one note spam and the next one help.\n\nIt sorts text into buckets for spam filters and help desks. It also helps teams spot angry posts online.", "humanExplainDisplay": "Text classification is a mailroom clerk\nwith a ==label maker==.\nIt stamps one note ==spam==\nand the next one help.\n\nIt sorts text into buckets\nfor spam filters and help desks.\nIt also helps teams\nspot angry posts online.", "relationsNarrative": "Classification\nText Classification is the text version of Classification.\n\nNLP\nText Classification is a common basic task in NLP.\n\nSupervised Learning\nText Classification often learns from labeled examples.\n\nBERT\nBERT is often used as the base model for Text Classification.", "relations": { "classification": { "label": "is the text version of …", "note": "It is Classification applied to text." }, "natural-language-processing": { "label": "belongs to …", "note": "It is one of the basic tasks in NLP." }, "supervised-learning": { "label": "often learns with …", "note": "Human labels teach the model how to sort text." }, "bert": { "label": "often uses …", "note": "BERT is often the base model for text classification." } } }, "zh": { "fullName": "Text Classification（文本分类）", "factExplain": "把文本自动分配到预设标签的 NLP 任务。", "humanExplain": "文本分类像居委会大妈看留言：报修、投诉、广告，扫一眼就贴条归队。\n\n用于邮件过滤、舆情和工单分流，让海量文字先归类。", "humanExplainDisplay": "文本分类像==居委会大妈==\n看留言：\n报修、投诉、广告，\n扫一眼就贴条归队。\n\n用于邮件过滤、舆情，\n和工单分流，\n让海量文字先归类。", "relationsNarrative": "Classification\n它是分类任务在文本数据上的具体应用。\n\nNLP\n它是自然语言处理里最常见的基础任务之一。\n\nSupervised Learning\n多数场景用带标签样本训练分类器。\n\nBERT\nBERT 常被微调用来做文本分类。", "relations": { "classification": { "label": "是…的文本版", "note": "它是分类任务在文本上的版本。" }, "natural-language-processing": { "label": "属于…", "note": "NLP 里最常见的基础任务之一。" }, "supervised-learning": { "label": "常用…训练", "note": "人工标签常用来教模型分门别类。" }, "bert": { "label": "用…建模", "note": "BERT 常作为文本分类的底座。" } } } } }, { "id": "text-summarization", "name": "Text Summary", "layer": "L4", "era": "1958", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "llm" }, { "to": "context-compression" }, { "to": "document-parsing" } ], "track": "history", "i18n": { "en": { "fullName": "Text Summarization", "factExplain": "Turning long text into a short version with the main points.", "humanExplain": "Text summarization is like a friend recapping a long movie. Skip the kissing subplot. Tell me who saved the planet.\n\nYou meet it in meeting notes, news, and document helpers. It saves time, but it can miss details.", "humanExplainDisplay": "Text summarization is like a friend\nrecapping a ==long movie==.\nSkip the kissing subplot.\nTell me ==who saved the planet==.\n\nYou meet it in meeting notes,\nnews, and document helpers.\nIt saves time,\nbut it can miss details.", "relationsNarrative": "NLP\nText summarization is a common text generation task in NLP.\n\nLLM\nAn LLM can read a long text and write a short summary.\n\nCtx Compression\nSummaries can squeeze long content into a shorter context.\n\nDocument parsing\nDocument parsing helps summaries keep the main points.", "relations": { "natural-language-processing": { "label": "is an … task", "note": "Summarization is a common text task in NLP." }, "llm": { "label": "is often written by …", "note": "An LLM can turn a long text into a human-like summary." }, "context-compression": { "label": "helps with …", "note": "A summary can squeeze long content into a shorter prompt." }, "document-parsing": { "label": "reads documents through …", "note": "Document parsing helps the summary see the document structure." } } }, "zh": { "fullName": "Text Summarization（文本摘要）", "factExplain": "把长文本压缩成保留要点的短文本任务。", "humanExplain": "文本摘要就是老板拍桌说讲重点：十页材料，压成三句能救会的干货。\n\n用于纪要、新闻和文档助手，省时但可能漏细节。", "humanExplainDisplay": "文本摘要就是老板拍桌\n说==讲重点==：\n十页材料，\n压成三句能救会的干货。\n\n用于纪要、新闻，\n和文档助手，\n省时但可能漏细节。", "relationsNarrative": "NLP\n文本摘要是 NLP 中典型的文本生成任务。\n\nLLM\nLLM 常用来理解长文并生成摘要。\n\nCtx Compression\n摘要可把长内容压缩进更短上下文。\n\nDocument parsing\n解析文档结构后，摘要更不容易漏重点。", "relations": { "natural-language-processing": { "label": "属于…任务", "note": "摘要是 NLP 里常见的文本任务。" }, "llm": { "label": "常由…生成", "note": "LLM 让摘要更像人工提炼。" }, "context-compression": { "label": "用于…", "note": "摘要可把长上下文压成短提示。" }, "document-parsing": { "label": "依赖…读文档", "note": "先解析文档，摘要才懂结构。" } } } } }, { "id": "text-to-image-generation", "name": "Text-to-Image Generation", "layer": "L4", "era": "2021", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "diffusion" }, { "to": "prompt" }, { "to": "clip" }, { "to": "deepfake" } ], "track": "history", "i18n": { "en": { "fullName": "Text-to-Image Generation", "factExplain": "An AI task that creates an image from a written description.", "humanExplain": "It is like a school art kid with rocket shoes. Say “a raccoon astronaut on a skateboard,” and boom, poster.\n\nPeople use it for posters and concept art. It makes rough drafts fast, but words and tiny details can fail.", "humanExplainDisplay": "It is like a ==school art kid==\nwith rocket shoes.\nSay ==“a raccoon astronaut on a skateboard,”==\nand boom, poster.\n\nPeople use it for posters and concept art.\nIt makes rough drafts fast,\nbut words and tiny details can fail.", "relationsNarrative": "Diffusion\nText-to-image generation is often built with Diffusion models.\n\nPrompt\nThe Prompt tells it what to draw and what style to use.\n\nCLIP\nCLIP helps match the written description to the image content.\n\nDeepfake\nText-to-image tools can also make fake images that mislead people.", "relations": { "diffusion": { "label": "is often built with …", "note": "Most modern text-to-image tools use Diffusion models." }, "prompt": { "label": "describes the image with …", "note": "The Prompt tells the model what the picture should look like." }, "clip": { "label": "aligns text and image with …", "note": "CLIP helps match words with image content." }, "deepfake": { "label": "can be used to make …", "note": "It can also make many realistic fake images." } } }, "zh": { "fullName": "Text-to-Image Generation｜文生图", "factExplain": "根据文本描述生成对应图像的生成任务。", "humanExplain": "你甩一句“赛博熊猫骑电驴”，它像夜市捏糖人师傅，听完就能当场给你捏个八九不离十。\n\n常用于海报、概念图和素材草稿；出图快，但文字与细节偶尔翻车。", "humanExplainDisplay": "你甩一句\n“赛博熊猫骑电驴”，\n它像==夜市捏糖人==师傅，\n听完就能当场给你\n捏个==八九不离十==。\n\n常用于海报、概念图\n和素材草稿；\n出图快，但文字与细节\n偶尔翻车。", "relationsNarrative": "Diffusion\n如今主流文生图，多由扩散模型来实现。\n\nPrompt\n提示词负责描述它要画什么、画成什么风格。\n\nCLIP\nCLIP 帮助模型把文字描述和图像内容对齐。\n\nDeepfake\n文生图也可能被用来生成具有误导性的伪造图片。", "relations": { "diffusion": { "label": "常由…实现", "note": "如今主流文生图大多基于扩散模型。" }, "prompt": { "label": "靠…描述画面", "note": "提示词决定构图、风格与主体细节。" }, "clip": { "label": "借…对齐图文", "note": "图文对齐能力帮助理解文字描述。" }, "deepfake": { "label": "可被用于生成…", "note": "它也可能被拿来批量伪造逼真图像。" } } } } }, { "id": "tf-idf", "name": "TF-IDF", "layer": "L2", "era": "1972", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "bag-of-words" }, { "to": "information-retrieval" }, { "to": "bm25" }, { "to": "vector-search" } ], "track": "history", "i18n": { "en": { "fullName": "Term Frequency-Inverse Document Frequency", "factExplain": "A text method for giving words scores based on frequency and rarity.", "humanExplain": "TF-IDF is like spotting toppings on pizza boxes. “Cheese” is everywhere, but one box yelling “anchovy” twice gets noticed.\n\nIn search, it pushes common words down. It lifts words with stronger topic clues.", "humanExplainDisplay": "TF-IDF is like spotting toppings on ==pizza boxes==.\n“Cheese” is everywhere,\nbut one box yelling ==“anchovy” twice== gets noticed.\n\nIn search,\nit pushes common words down.\nIt lifts words with stronger topic clues.", "relationsNarrative": "Bag-of-Words\nTF-IDF starts with Bag-of-Words, then gives each word a weight.\n\nIR\nTF-IDF is a classic signal for search ranking.\n\nBM25\nBM25 is a stronger version built for search ranking.\n\nVector search\nTF-IDF leans on keyword matches. Vector search leans on meaning.", "relations": { "bag-of-words": { "label": "weights …", "note": "TF-IDF adds word importance scores on top of Bag-of-Words." }, "information-retrieval": { "label": "supports … ranking", "note": "Classic search uses TF-IDF to score text relevance." }, "bm25": { "label": "inspired …", "note": "BM25 is a search-friendly upgrade from the same idea." }, "vector-search": { "label": "contrasts with …", "note": "TF-IDF follows keywords, while vectors follow meaning." } } }, "zh": { "fullName": "词频-逆文档频率", "factExplain": "按词频和稀有度给词加权的文本表示方法。", "humanExplain": "TF-IDF像班主任查考勤：天天喊到不稀罕，稀有名字一出现就划重点。\n\n搜索排序时压低常见词，突出更能区分主题的词。", "humanExplainDisplay": "TF-IDF像\n==班主任查考勤==：\n天天喊到不稀罕，\n稀有名字一出现就==划重点==。\n\n搜索排序时压低常见词，\n突出更能区分主题的词。", "relationsNarrative": "Bag-of-Words\nTF-IDF 在词袋基础上，为每个词加权。\n\nInformation Retrieval\nTF-IDF 是传统搜索排序的经典信号。\n\nBM25\nBM25 可看作更适合搜索排序的改进版。\n\nVector Search\n它偏关键词匹配，向量搜索偏语义相似。", "relations": { "bag-of-words": { "label": "给…加权", "note": "在词袋基础上衡量词的重要性。" }, "information-retrieval": { "label": "支撑…排序", "note": "传统搜索常用它计算文本相关性。" }, "bm25": { "label": "启发…改进", "note": "BM25 可看作搜索版升级思路。" }, "vector-search": { "label": "对比…路线", "note": "它看关键词，向量更看语义。" } } } } }, { "id": "the-bitter-lesson", "name": "The Bitter Lesson", "layer": "L1", "era": "2019", "publishedAt": "2026-06-27T04:00:00.000Z", "relations": [ { "to": "scaling-law" }, { "to": "deep-learning" }, { "to": "symbolic-ai" }, { "to": "feature-engineering" } ], "track": "history", "i18n": { "en": { "fullName": "The Bitter Lesson", "factExplain": "An AI lesson: general computing often beats hand-written expert rules.", "humanExplain": "The Bitter Lesson is a chess match. A master's clever tricks look smarter. But once compute is cheap, a machine that just searches more wins anyway.\n\nIn AI, general methods win as data and compute grow. It helps you pick the better AI bet.", "humanExplainDisplay": "The Bitter Lesson is a chess match.\nA master's ==clever tricks look smarter==.\nBut once compute is cheap,\na machine that ==just searches more wins anyway==.\n\nIn AI, general methods win\nas data and compute grow.\nIt helps you pick the better AI bet.", "relationsNarrative": "Scaling-law\nThe Bitter Lesson explains why scalable computing keeps getting stronger.\n\nDeep Learning\nDeep Learning follows this path: fewer rules, more data, and more compute.\n\nSymbolic AI\nThe Bitter Lesson warns that hand-written knowledge can lose to scalable learning.\n\nFeature-engineering\nThe Bitter Lesson turns hand-made features from the star into a helper.", "relations": { "scaling-law": { "label": "explains why … works", "note": "More compute helps general methods keep getting better." }, "deep-learning": { "label": "backs the rise of …", "note": "Deep Learning writes fewer rules and learns from data." }, "symbolic-ai": { "label": "warns against overtrusting …", "note": "Hand-written knowledge can be smart but hard to scale." }, "feature-engineering": { "label": "shrinks the role of …", "note": "When features can be learned, hand-made ones matter less." } } }, "zh": { "fullName": "苦涩的教训", "factExplain": "一条主张通用计算最终胜过手工知识的 AI 经验。", "humanExplain": "苦涩教训像下棋：老师傅的妙招看着更聪明，可一旦算力管够，只会硬算的电脑照样赢。\n\n教 AI 别死磕人类技巧，押可扩展的数据和算力。", "humanExplainDisplay": "苦涩教训像下棋：\n老师傅的==妙招看着更聪明==，\n可一旦算力管够，\n只会==硬算的电脑照样赢==。\n\n教 AI 别死磕人类技巧，\n押可扩展的数据和算力。", "relationsNarrative": "Scaling-law\n苦涩教训解释了为什么可扩展计算会越跑越强。\n\nDeep Learning\n深度学习正是少写规则、多靠数据算力的路线。\n\nSymbolic AI\n它提醒人类手写知识常会输给可扩展学习。\n\nFeature-engineering\n它让手搓特征从主角变成辅助。", "relations": { "scaling-law": { "label": "解释…为何有效", "note": "算力增加时，通用方法持续吃香。" }, "deep-learning": { "label": "支撑…的胜利", "note": "它少写规则，多靠数据自动学习。" }, "symbolic-ai": { "label": "提醒别迷信…", "note": "手工知识很聪明，但不一定可扩展。" }, "feature-engineering": { "label": "削弱…的地位", "note": "特征越能自动学，手搓越不值钱。" } } } } }, { "id": "third-party-ai-evaluation", "name": "Third-party AI evaluation", "layer": "L4", "era": "2024", "publishedAt": "2026-05-30T03:10:23.229Z", "relations": [ { "to": "model-leaderboard" }, { "to": "ai-regulation" }, { "to": "llm-as-a-judge" }, { "to": "benchmark-contamination" } ], "track": "today", "i18n": { "en": { "fullName": "Third-party AI evaluation", "factExplain": "An outside group tests an AI system and judges how well it works.", "humanExplain": "The seller says the roof is perfect. You still want a home inspector with a ladder.\n\nAn outside group tests the AI. It helps before launch. It helps buyers pick. It helps regulators judge risk.", "humanExplainDisplay": "The seller says the roof is perfect.\nYou still want ==a home inspector==\nwith ==a ladder==.\n\nAn outside group tests the AI.\nIt helps before launch.\nIt helps buyers pick.\nIt helps regulators judge risk.", "relationsNarrative": "Leaderboard\nMany Leaderboard scores are built on third-party eval results.\n\nAI-regulation\nThird-party eval can give AI-regulation a more independent risk view.\n\nLLM-as-a-judge\nSome third-party evals use LLM-as-a-judge to help score answers.\n\nBenchmark contamination\nBenchmark contamination can make a third-party eval score look too high.", "relations": { "model-leaderboard": { "label": "supports …", "note": "Many leaderboard scores come from outside tests." }, "ai-regulation": { "label": "guides …", "note": "Regulators use evals to judge AI risk." }, "llm-as-a-judge": { "label": "may use … to score", "note": "Some outside evals use an LLM to help score answers." }, "benchmark-contamination": { "label": "must guard against …", "note": "Leaked test questions can make scores lie." } } }, "zh": { "fullName": "第三方 AI 评测", "factExplain": "由独立机构对 AI 系统进行外部测试和评估。", "humanExplain": "第三方评测像相亲带闺蜜把关，别只听本人把自己夸成抢手货。\n\n它用于模型选型、榜单和风控，帮你少被宣传话术带跑。", "humanExplainDisplay": "第三方评测像==相亲带闺蜜把关==，\n别只听本人把自己夸成==抢手货==。\n\n它用于模型选型、榜单和风控，\n帮你少被宣传话术带跑。", "relationsNarrative": "Model-leaderboard\n许多模型排行榜的分数，都建立在第三方评测结果之上。\n\nAI-regulation\n第三方评测能为监管提供相对独立的风险判断依据。\n\nLLM-as-a-judge\n有些第三方评测会使用 LLM-as-a-judge 辅助评分。\n\nBenchmark contamination\n如果发生 Benchmark contamination，第三方评测也可能被刷高分数。", "relations": { "model-leaderboard": { "label": "为…提供依据", "note": "许多排行榜成绩来自外部评测。" }, "ai-regulation": { "label": "为…提供参考", "note": "监管常借评测判断模型风险水平。" }, "llm-as-a-judge": { "label": "可借助…打分", "note": "部分外部评测会让模型辅助评分。" }, "benchmark-contamination": { "label": "要防范…干扰", "note": "题目泄露会让评测结果失真。" } } } } }, { "id": "thompson-sampling", "name": "Thompson Sampling", "layer": "L2", "era": "1933", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "exploration-exploitation" }, { "to": "reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Thompson Sampling", "factExplain": "A decision method using uncertainty samples to balance new tries and likely winners.", "humanExplain": "Imagine three arcade claw machines. Try each one, then feed more quarters to the less cursed one.\n\nIt keeps testing while favoring the current favorite. Ads and recommendation feeds use it to send more traffic to what works.", "humanExplainDisplay": "Imagine ==three arcade claw machines==.\nTry each one,\nthen feed more quarters to\n==the less cursed one==.\n\nIt keeps testing while favoring the current favorite.\nAds and recommendation feeds use it\nto send more traffic to what works.", "relationsNarrative": "Exploration-Exploitation Tradeoff\nThompson Sampling balances trying new choices with using likely winners.\n\nRL\nThompson Sampling can guide online decisions as the system learns by acting.", "relations": { "exploration-exploitation": { "label": "handles …", "note": "It is a classic way to balance trying and using." }, "reinforcement-learning": { "label": "guides … decisions", "note": "It often guides online decisions in RL." } } }, "zh": { "fullName": "汤普森采样（Thompson Sampling）", "factExplain": "一种按不确定性采样来平衡探索与利用的决策方法。", "humanExplain": "夜市套圈别老盯着同一个奖品猛砸，先按手感挑几个都试试；哪个看着更有戏，就多分它几枚圈。\n\n常用于推荐、广告等在线分配，边试边学，把流量更多给更可能有效的选项。", "humanExplainDisplay": "夜市套圈别老盯着\n==同一个奖品猛砸==，\n先按手感挑几个都试试；\n哪个==看着更有戏==，\n就多分它几枚圈。\n\n常用于推荐、广告等\n在线分配，\n边试边学，\n把流量更多给更可能有效的选项。", "relationsNarrative": "Exploration-Exploitation Tradeoff\n它用概率采样来平衡试新选择和吃老本。\n\nReinforcement Learning\n它常作为在线决策策略，用于边行动边学习。", "relations": { "exploration-exploitation": { "label": "应对…权衡", "note": "它是经典的探索与利用平衡方法。" }, "reinforcement-learning": { "label": "常用于…决策", "note": "常见于在线学习和强化学习决策。" } } } } }, { "id": "token-efficiency", "name": "Token Efficiency", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "context-compression" }, { "to": "token-tax" }, { "to": "prompt-engineering" } ], "track": "today", "i18n": { "en": { "fullName": "Token Efficiency", "factExplain": "Doing the same AI task with fewer tokens.", "humanExplain": "Token efficiency is packing for a trip with one carry-on. You bring socks, not a waffle maker.\n\nIt matters in prompts and long documents. It saves money, time, and context space.", "humanExplainDisplay": "Token efficiency is ==packing for a trip==\nwith ==one carry-on==.\nYou bring socks,\nnot a waffle maker.\n\nIt matters in prompts\nand long documents.\nIt saves money, time,\nand context space.", "relationsNarrative": "Token\nToken efficiency checks if each Token is worth it.\n\nCtx Compression\nCtx Compression keeps the useful parts and spends fewer tokens.\n\nToken Tax\nLow token efficiency means a higher Token Tax.\n\nPrompt-engineering\nGood Prompt-engineering helps the model take a shorter path.", "relations": { "token": { "label": "measures value per …", "note": "Each Token costs money and uses context space." }, "context-compression": { "label": "uses … to cut waste", "note": "Ctx Compression keeps key info and spends fewer tokens." }, "token-tax": { "label": "lowers …", "note": "Extra words raise the Token Tax." }, "prompt-engineering": { "label": "improves through …", "note": "Good prompts take fewer wrong turns and use fewer tokens." } } }, "zh": { "fullName": "令牌效率", "factExplain": "用更少 token 完成同等任务的能力。", "humanExplain": "令牌效率像出门只带一个登机箱：袜子带上，华夫饼机留下，旅行照样完整。\n\n用于提示优化和长文处理，省钱省时，也省上下文。", "humanExplainDisplay": "令牌效率像出门\n只带==一个登机箱==：\n袜子带上，\n==华夫饼机留下==，\n旅行照样完整。\n\n用于提示优化\n和长文处理，\n省钱省时，也省上下文。", "relationsNarrative": "Token\n令牌效率衡量每个 Token 是否花得值。\n\nCtx Compression\n上下文压缩是提升令牌效率的常见办法。\n\nToken Tax\n令牌效率越低，额外的 Token 税越高。\n\nPrompt-engineering\n好提示能让模型少绕路、少消耗。", "relations": { "token": { "label": "衡量…用得值不值", "note": "每个 Token 都会占成本和上下文。" }, "context-compression": { "label": "靠…减少浪费", "note": "压缩上下文能保留信息、少花 Token。" }, "token-tax": { "label": "降低…", "note": "废话越多，Token 税越高。" }, "prompt-engineering": { "label": "用…提升", "note": "好提示能少绕路、少消耗。" } } } } }, { "id": "token-tax", "name": "Token Tax", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "context-window" }, { "to": "ai-finops" }, { "to": "cost-aware-ai-ai" } ], "track": "today", "i18n": { "en": { "fullName": "Token Tax", "factExplain": "The extra cost and delay caused by using more tokens.", "humanExplain": "Token Tax is like ordering pizza by the topping. Add extra olives, and the bill grows while you wait.\n\nYou meet it in agents and support bots. Long document Q&A feels it too. It slows replies and raises costs.", "humanExplainDisplay": "Token Tax is like ordering pizza ==by the topping==.\nAdd extra olives,\nand the ==bill grows== while you wait.\n\nYou meet it in agents and support bots.\nLong document Q&A feels it too.\nIt slows replies and raises costs.", "relationsNarrative": "Token\nToken Tax turns each extra Token into cost.\n\nContext-window\nA longer Context-window can pile the tax higher.\n\nAI FinOps\nAI FinOps tracks and cuts this hidden cost.\n\nCost-aware AI\nCost-aware AI spends fewer tokens on purpose.", "relations": { "token": { "label": "adds cost per …", "note": "Each Token can become part of the bill." }, "context-window": { "label": "grows with …", "note": "A longer Context-window can hide a bigger cost." }, "ai-finops": { "label": "is tracked by …", "note": "AI FinOps watches this hidden cost and cuts it down." }, "cost-aware-ai-ai": { "label": "pushes … design", "note": "Cost-aware AI starts by wasting fewer tokens." } } }, "zh": { "fullName": "令牌税", "factExplain": "额外 token 带来的成本与延迟。", "humanExplain": "令牌税像高速过路费：车没开多远，收费站一过，账单就冒头。\n\n常见于智能体、客服、长文问答，拖慢响应并抬高成本。", "humanExplainDisplay": "令牌税像==高速过路费==：\n车没开多远，\n收费站一过，\n==账单就冒头==。\n\n常见于智能体、客服、长文问答，\n拖慢响应，\n并抬高成本。", "relationsNarrative": "Token\n它把每个额外 Token 都变成成本。\n\nContext-window\n上下文越长，越容易把税堆高。\n\nAI FinOps\nAI FinOps 会追踪并削减这类隐性成本。\n\nCost-aware AI\n成本敏感设计会主动少花这些 token。", "relations": { "token": { "label": "按…累积成本", "note": "每个 token 都可能变成账单。" }, "context-window": { "label": "随…放大开销", "note": "上下文越长，隐性开销越高。" }, "ai-finops": { "label": "被…追踪优化", "note": "它是 AI FinOps 常抓的成本项。" }, "cost-aware-ai-ai": { "label": "倒逼…设计", "note": "省 token 是成本意识的第一课。" } } } } }, { "id": "token", "name": "Token", "layer": "L2", "era": "2010s", "publishedAt": "2026-05-23T08:05:00Z", "relations": [ { "to": "llm" }, { "to": "context-window" }, { "to": "inference" } ], "i18n": { "en": { "fullName": "Token", "factExplain": "A token is the small text unit AI reads and writes.", "humanExplain": "AI does not swallow a sentence whole. It eats text in little nuggets called tokens.\n\nTokens set how much text a model can read at once. More tokens also mean a bigger API bill.", "humanExplainDisplay": "AI does not swallow ==a sentence whole==.\nIt eats text in ==little nuggets== called tokens.\n\nTokens set how much text\na model can read at once.\nMore tokens also mean\na bigger API bill.", "relationsNarrative": "LLM\nA token is the basic text unit an LLM uses for language.\n\nContext-window\nA Context-window can hold only a set number of tokens.\n\nInference\nInference costs more when input and output tokens increase.", "relations": { "llm": { "label": "is the text unit for …", "note": "A token is the basic text unit an LLM works with." }, "context-window": { "label": "measures length in …", "note": "A context window is counted by how many tokens it can hold." }, "inference": { "label": "gets used during …", "note": "Inference costs rise as input and output tokens increase." } } }, "zh": { "fullName": "词元", "factExplain": "AI 读取和生成文本时使用的最小计量单位。", "humanExplain": "Token 像 AI 吃文字时切出来的小块肉，句子越长，切得越多。\n\n它决定模型能看多少内容，也决定你调用一次 API 要烧多少钱。", "humanExplainDisplay": "Token 像 AI 吃文字时\n掉下来的==小碎块==。\n一句话越长，切得越碎。\n\n它决定模型能看多少内容，\n也决定你调用一次 API\n到底有多烧字。", "relationsNarrative": "LLM\nToken 是 LLM 处理语言时最基本的计算单位。\n\nContext-window\nContext-window 能装下多少内容，取决于 Token 容量。\n\nInference\nInference 的成本会随着输入和输出 Token 增加。", "relations": { "llm": { "label": "是…的计量单位" }, "context-window": { "label": "按…计长度" }, "inference": { "label": "在…时被消耗" } } } } }, { "id": "tokens-per-second", "name": "TPS", "layer": "L5", "sublayer": "compute", "era": "2020s", "publishedAt": "2026-06-01T04:00:00.000Z", "relations": [ { "to": "token" }, { "to": "inference" }, { "to": "continuous-batching" }, { "to": "gpu" } ], "track": "today", "i18n": { "en": { "fullName": "Tokens per second", "factExplain": "A speed score for how fast an AI writes during inference.", "humanExplain": "TPS is the AI’s words-per-minute score. A fast bot sounds like an auctioneer. A slow one types like your uncle with one finger.\n\nIt shapes chat wait time and autocomplete smoothness. People use it to compare inference services.", "humanExplainDisplay": "TPS is the AI’s ==words-per-minute score==.\nA fast bot sounds like ==an auctioneer==.\nA slow one types like your uncle\nwith one finger.\n\nIt shapes chat wait time\nand autocomplete smoothness.\nPeople use it\nto compare inference services.", "relationsNarrative": "Token\nTPS counts tokens made each second.\n\nInference\nTPS measures how fast the model generates during inference.\n\nContinuous batching\nContinuous batching cuts idle time, so TPS often goes up.\n\nGPU\nGPU power and memory speed can limit TPS.", "relations": { "token": { "label": "counts in …", "note": "TPS counts how many tokens the model makes each second." }, "inference": { "label": "measures … speed", "note": "TPS is a common speed score during inference." }, "continuous-batching": { "label": "is boosted by …", "note": "Better scheduling often raises TPS." }, "gpu": { "label": "depends on … power", "note": "Stronger GPUs usually let the model generate faster." } } }, "zh": { "fullName": "Tokens per second（每秒生成 token 数）", "factExplain": "衡量模型推理生成速度的常用指标。", "humanExplain": "同样一道题，有的 AI 输出快得像==嘴皮子开挂==，有的还在字斟句酌；TPS 量的就是这口条速度。\n\n它直接影响聊天等待感和补全流畅度，也常拿来比较推理服务性能。", "humanExplainDisplay": "同样一道题，\n有的 AI 输出快得像 \n==嘴皮子开挂==，\n有的还在字斟句酌；\nTPS 量的就是这 \n==口条速度==。\n\n它直接影响聊天等待感\n和补全流畅度，\n也常拿来比较推理服务性能。", "relationsNarrative": "Token\nTPS 的计数单位就是 token，数的是每秒吐出多少个。\n\nInference\nTPS 主要用来衡量模型在推理阶段的生成速度。\n\nContinuous batching\nContinuous batching 能减少空转，常让 TPS 表现更高。\n\nGPU\nGPU 算力和带宽会直接影响 TPS 上限。", "relations": { "token": { "label": "以…为计数单位", "note": "TPS 数的就是每秒生成多少 token。" }, "inference": { "label": "衡量…速度", "note": "它是推理阶段最常见的速度指标之一。" }, "continuous-batching": { "label": "被…影响表现", "note": "调度更高效时，TPS 往往也会提升。" }, "gpu": { "label": "受…算力限制", "note": "显卡算力越强，通常生成速度越高。" } } } } }, { "id": "tpu", "name": "TPU", "layer": "L5", "sublayer": "compute", "era": "2016", "publishedAt": "2026-06-21T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "ai-chip" }, { "to": "ai-data-center" }, { "to": "compute-race" } ], "track": "history", "i18n": { "en": { "fullName": "Tensor Processing Unit", "factExplain": "A special AI chip built to run machine learning math fast.", "humanExplain": "A TPU is like the fry cook at a packed diner. It does one job, but the fries fly out by the bucket.\n\nIt runs big AI training and AI answers. You usually meet it in cloud data centers. It is built to be fast. It saves power. It works in huge packs.", "humanExplainDisplay": "A TPU is like the ==fry cook==\nat a packed diner.\nIt does ==one job==,\nbut the fries fly out by the bucket.\n\nIt runs big AI training and AI answers.\nYou usually meet it in cloud data centers.\nIt is built to be fast.\nIt saves power.\nIt works in huge packs.", "relationsNarrative": "GPU\nBoth can run AI work, but a TPU is more specialized.\n\nAI chip\nA TPU is an AI chip built for machine learning.\n\nAI data center\nTPUs often sit in data centers for large training and inference jobs.\n\nCompute-race\nStronger TPUs make the compute race heat up.", "relations": { "gpu": { "label": "splits AI work with …", "note": "Both run AI, but their designs lean different ways." }, "ai-chip": { "label": "is a kind of …", "note": "A TPU is a special chip built for AI math." }, "ai-data-center": { "label": "runs inside …", "note": "TPUs are often used in large groups for big training jobs." }, "compute-race": { "label": "pushes … higher", "note": "Stronger TPUs raise the stakes in the compute race." } } }, "zh": { "fullName": "Tensor Processing Unit｜张量处理单元", "factExplain": "一种为机器学习计算专门设计的 AI 芯片。", "humanExplain": "TPU 不像瑞士军刀，更像工地上的钉枪：活儿单一，但钉子一排排打出去，快得让人眼皮直跳。\n\n主要跑大模型训练和推理，常见于云端集群，强调速度、能效和规模。", "humanExplainDisplay": "TPU 不像瑞士军刀，\n更像工地上的\n==钉枪==：\n活儿单一，\n但钉子一排排\n==打出去==，\n快得让人眼皮直跳。\n\n主要跑大模型训练和推理，\n常见于云端集群，\n强调速度、能效\n和规模。", "relationsNarrative": "GPU\n两者都能跑 AI 计算，但 TPU 更偏专用化设计。\n\nAI chip\nTPU 属于 AI 芯片，是为机器学习优化的硬件。\n\nAI data center\nTPU 常部署在数据中心，支撑大规模训练与推理。\n\nCompute-race\n更强的 TPU 会加码算力竞争，影响行业节奏。", "relations": { "gpu": { "label": "对比…分工", "note": "两者都跑 AI，但设计取向不同。" }, "ai-chip": { "label": "属于…一类", "note": "它是面向 AI 计算的专用芯片。" }, "ai-data-center": { "label": "部署在…里", "note": "常成批部署支撑大规模训练。" }, "compute-race": { "label": "推动…升级", "note": "更强专用芯片会抬高算力竞争。" } } } } }, { "id": "transfer-learning", "name": "Transfer Learning", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "pretraining" }, { "to": "fine-tuning" }, { "to": "foundation-model" }, { "to": "bert" } ], "track": "history", "i18n": { "en": { "fullName": "Transfer Learning", "factExplain": "A way to reuse what one model learned for a new task.", "humanExplain": "Transfer learning is like a soccer kid trying basketball. The ball changed, but the footwork still helps.\n\nTeams often start with a pretrained model. A little new data can teach it to sort tickets, read photos, or learn company tasks.", "humanExplainDisplay": "Transfer learning is like\na ==soccer kid== trying basketball.\nThe ball changed,\nbut the ==footwork still helps==.\n\nTeams often start with a pretrained model.\nA little new data can teach it to sort tickets,\nread photos,\nor learn company tasks.", "relationsNarrative": "Pretraining\nTransfer learning often starts with general skills from pretraining.\n\nFine-tuning\nFine-tuning is a common way to move old skills to a new task.\n\nFoundation-model\nA stronger Foundation-model often needs less new data for the next task.\n\nBERT\nBERT helped make pretraining plus transfer learning normal in NLP.", "relations": { "pretraining": { "label": "builds on …", "note": "It learns general skills first, then moves them to a task." }, "fine-tuning": { "label": "often lands through …", "note": "Fine-tuning is a common way to adapt old skills." }, "foundation-model": { "label": "often starts from …", "note": "A stronger foundation model needs less new data for transfer." }, "bert": { "label": "became common with …", "note": "BERT made pretrain-then-transfer normal in NLP." } } }, "zh": { "fullName": "迁移学习", "factExplain": "把已有模型学到的能力迁到新任务上的方法。", "humanExplain": "迁移学习像厨师换摊卖煎饼：火候、手感、刀工都还在，不用从生火学起，换个配方就能很快出餐。\n\n它常以预训练模型为底座，用少量新数据完成分类、识图和行业微调。", "humanExplainDisplay": "迁移学习像厨师换摊卖煎饼：\n火候、手感、刀工都还在，\n不用从==生火学起==，\n换个配方就能==很快出餐==。\n\n它常以预训练模型为底座，\n用少量新数据完成分类、\n识图和行业微调。", "relationsNarrative": "Pretraining\n迁移学习通常先承接预训练得到的通用能力。\n\nFine-tuning\n微调是把已有能力迁到新任务的常见做法。\n\nFoundation-model\n基础模型越强，迁移到下游任务通常越省数据。\n\nBERT\nBERT 让预训练加迁移学习成为 NLP 主流。", "relations": { "pretraining": { "label": "承接…成果", "note": "先学通用能力，再迁到具体任务。" }, "fine-tuning": { "label": "常靠…落地", "note": "微调是迁移到新任务的常见做法。" }, "foundation-model": { "label": "常以…为底座", "note": "大模型让迁移学习更通用更省数据。" }, "bert": { "label": "在…中常见", "note": "BERT 带火了预训练后迁移的范式。" } } } } }, { "id": "transformer", "name": "Transformer", "layer": "L2", "era": "2017", "publishedAt": "2026-05-23T08:50:00Z", "relations": [ { "to": "attention" }, { "to": "llm" }, { "to": "neural-network" }, { "to": "foundation-model" } ], "i18n": { "en": { "fullName": "Transformer architecture", "factExplain": "A deep learning design that uses attention to understand sequences.", "humanExplain": "A Transformer is the kid at a loud lunch table. It hears “pass the ketchup” and knows the fries are involved.\n\nIt helps AI connect words in context and write text. Many LLMs and foundation models use it as their main engine.", "humanExplainDisplay": "A Transformer is the ==kid at a loud lunch table==.\nIt hears “pass the ketchup”\nand knows the ==fries are involved==.\n\nIt helps AI connect words in context\nand write text.\nMany LLMs and foundation models\nuse it as their main engine.", "relationsNarrative": "Attention\nAttention helps Transformer catch links across the context.\n\nLLM\nTransformer helped modern LLMs make a big jump in skill.\n\nNeural-network\nTransformer is a Neural-network design built for sequence tasks.\n\nFoundation-model\nMany Foundation-models use Transformer as their general design.", "relations": { "attention": { "label": "runs on …", "note": "Attention is the core tool Transformer uses to track context." }, "llm": { "label": "powers …", "note": "Transformer made modern LLMs much better at language." }, "neural-network": { "label": "is a kind of …", "note": "Transformer is a Neural-network design for sequence tasks." }, "foundation-model": { "label": "supports …", "note": "Many Foundation-models use Transformer as their general design." } } }, "zh": { "fullName": "Transformer 架构", "factExplain": "一种依靠注意力机制处理序列信息的深度学习架构。", "humanExplain": "Transformer 像开会时特别会抓重点的人，满屋子发言它能听出谁和谁有关。\n\n今天很多 LLM 都靠它理解上下文、生成文本，是 AI 爆发的关键底座。", "humanExplainDisplay": "Transformer 像开会时\n==特别会抓重点的人==。\n一屋子发言，它能听出谁和谁有关。\n\n很多 LLM 都靠它理解上下文。\nAI 这波爆发，\n它算幕后大功臣。", "relationsNarrative": "Attention\nAttention 是 Transformer 捕捉上下文关系的核心机制。\n\nLLM\nTransformer 架构推动了现代 LLM 的能力跃迁。\n\nNeural-network\nTransformer 是 Neural-network 中适合序列建模的结构。\n\nFoundation-model\n许多 Foundation-model 以 Transformer 作为通用架构。", "relations": { "attention": { "label": "靠…机制运转" }, "llm": { "label": "是…的基础架构" }, "neural-network": { "label": "是…的一种" }, "foundation-model": { "label": "支撑…" } } } } }, { "id": "trec", "name": "TREC", "layer": "L4", "era": "1992", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "information-retrieval" }, { "to": "bm25" }, { "to": "question-answering" } ], "track": "history", "i18n": { "en": { "fullName": "Text REtrieval Conference", "factExplain": "A NIST-run conference for testing how well search systems retrieve text.", "humanExplain": "TREC is a school quiz bowl for search engines. Everyone gets the same questions, so no one brags about an easy sheet.\n\nIt gives standard test sets for retrieval. People use them to judge search. They also use them for QA and RAG.", "humanExplainDisplay": "TREC is a ==school quiz bowl==\nfor search engines.\nEveryone gets the ==same questions==,\nso no one brags about an easy sheet.\n\nIt gives standard test sets for retrieval.\nPeople use them to judge search.\nThey also use them for QA and RAG.", "relationsNarrative": "IR\nTREC gives IR shared tasks and data for fair testing.\n\nBM25\nBM25 is often used as a baseline on TREC data.\n\nQA\nTREC's QA track helped shape open-domain QA testing.", "relations": { "information-retrieval": { "label": "sets exams for …", "note": "Shared tasks make retrieval systems easy to compare." }, "bm25": { "label": "often uses … as a baseline", "note": "BM25 is often scored on TREC datasets." }, "question-answering": { "label": "opened a track for …", "note": "TREC helped push open-domain QA testing forward." } } }, "zh": { "fullName": "Text REtrieval Conference，文本检索会议", "factExplain": "由 NIST 主办的检索评测会议。", "humanExplain": "TREC像给搜索引擎办统考：人人做同一张卷，分数才比得出真高低。\n\n提供检索评测基准，用于搜索、问答和RAG评估。", "humanExplainDisplay": "TREC像给搜索引擎\n办==统考==：\n人人做==同一张卷==，\n分数才比得出真高低。\n\n提供检索评测基准，\n用于搜索、问答\n和RAG评估。", "relationsNarrative": "Information Retrieval\nTREC 提供统一任务和数据，推动检索评测成型。\n\nBM25\nBM25 常被拿来在 TREC 数据上做基线对比。\n\nQuestion Answering\nTREC 的问答赛道推动了开放域问答评测。", "relations": { "information-retrieval": { "label": "给…出考卷", "note": "统一任务让检索系统可比较。" }, "bm25": { "label": "常用…作基线", "note": "BM25 常在 TREC 数据上跑分。" }, "question-answering": { "label": "开辟…赛道", "note": "QA 赛道推动开放域问答评测。" } } } } }, { "id": "trillion-parameter-model", "name": "Trillion-parameter model", "layer": "L3", "era": "2021", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "scaling-law" }, { "to": "foundation-model" }, { "to": "model-parallelism" } ], "track": "today", "i18n": { "en": { "fullName": "Trillion-parameter model", "factExplain": "A huge AI model with about one trillion learned settings.", "humanExplain": "A trillion-parameter model is a stadium-sized kitchen. It can cook almost anything, but the power bill cries.\n\nYou meet it behind top AI assistants and serious coding help. Science labs use it too, but training and running it costs a fortune.", "humanExplainDisplay": "A trillion-parameter model is a ==stadium-sized kitchen==.\nIt can cook almost anything,\nbut the ==power bill cries==.\n\nYou meet it behind top AI assistants\nand serious coding help.\nScience labs use it too,\nbut training and running it costs a fortune.", "relationsNarrative": "Parameter\nParameter count is the core reason for this model's name.\n\nScaling-law\nScaling-law gives people a reason to make models bigger.\n\nFoundation-model\nThis model often becomes a general base for many tasks.\n\nModel parallelism\nModel parallelism splits the huge model across many chips for training and use.", "relations": { "parameter": { "label": "gets its name from …", "note": "The parameter count gives this model its name." }, "scaling-law": { "label": "is inspired by …", "note": "Scaling-law says bigger models often get stronger." }, "foundation-model": { "label": "often serves as …", "note": "It is often trained as a general base for many tasks." }, "model-parallelism": { "label": "trains with …", "note": "The model is too big for one chip, so many chips share it." } } }, "zh": { "fullName": "万亿参数模型", "factExplain": "参数规模约达一万亿的超大模型。", "humanExplain": "万亿参数模型像航空母舰：吨位巨大、火力全开，可一开动就烧天文数字的油钱。\n\n用于顶级助手、代码和科研，训练部署门槛极高。", "humanExplainDisplay": "万亿参数模型像\n==航空母舰==：\n吨位巨大、火力全开，\n可一开动就烧\n==天文数字的油钱==。\n\n用于顶级助手、代码和科研，\n训练部署门槛极高。", "relationsNarrative": "Parameter\n参数数量是万亿参数模型的命名核心。\n\nScaling-law\nScaling-law 支撑了把模型做大的基本预期。\n\nFoundation-model\n这类模型常被训练成多任务通用底座。\n\nModel parallelism\n模型太大时，需要切分到多卡训练和推理。", "relations": { "parameter": { "label": "堆大…规模", "note": "参数数量是这类模型的命名核心。" }, "scaling-law": { "label": "受…启发", "note": "规模变大常带来能力提升预期。" }, "foundation-model": { "label": "常作为…", "note": "它通常服务多任务通用能力。" }, "model-parallelism": { "label": "依赖…训练", "note": "单卡放不下，只能拆开协作。" } } } } }, { "id": "trust-region-policy-optimization", "name": "TRPO", "layer": "L2", "era": "2015", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "reinforcement-learning" }, { "to": "policy-gradient" }, { "to": "proximal-policy-optimization" }, { "to": "kullback-leibler-divergence" } ], "track": "history", "i18n": { "en": { "fullName": "Trust Region Policy Optimization", "factExplain": "An RL algorithm that limits how far a policy can change at once.", "humanExplain": "TRPO is bowling with bumpers for a learning robot. It can roll better, but not launch the ball into lane 6.\n\nIt makes policy updates safer. You meet it in RL work for robots and game control.", "humanExplainDisplay": "TRPO is ==bowling with bumpers==\nfor a learning robot.\nIt can roll better,\nbut not ==launch the ball into lane 6==.\n\nIt makes policy updates safer.\nYou meet it in RL work\nfor robots and game control.", "relationsNarrative": "RL\nTRPO is an RL method for updating a policy safely.\n\nPolicy Gradient\nTRPO adds a safety boundary to Policy Gradient updates.\n\nPPO\nPPO is a simpler version inspired by TRPO.\n\nKL Divergence\nTRPO uses KL Divergence to keep the new policy close to the old one.", "relations": { "reinforcement-learning": { "label": "belongs to …", "note": "It is an RL method for safer policy updates." }, "policy-gradient": { "label": "adds guardrails to …", "note": "It puts a step limit on Policy Gradient updates." }, "proximal-policy-optimization": { "label": "inspired …", "note": "PPO copies the idea in a simpler way." }, "kullback-leibler-divergence": { "label": "limits steps with …", "note": "KL Divergence measures how far the new policy moved." } } }, "zh": { "fullName": "Trust Region Policy Optimization（信赖域策略优化）", "factExplain": "用信赖域限制策略更新幅度的强化学习算法。", "humanExplain": "TRPO像驾校教练踩副刹：允许你进步，方向盘别一把打进沟。\n\n让策略更新更稳，常用于机器人和游戏控制。", "humanExplainDisplay": "TRPO像驾校教练==踩副刹==：\n允许你进步，\n方向盘别==一把打进沟==。\n\n让策略更新更稳，\n常用于机器人\n和游戏控制。", "relationsNarrative": "Reinforcement Learning\nTRPO 是强化学习里更新策略的训练算法。\n\nPolicy Gradient\n它在策略梯度更新上加了安全边界。\n\nPPO\nPPO 是它的简化版，更易实现和调参。\n\nKL Divergence\n它用 KL 约束新旧策略别差太远。", "relations": { "reinforcement-learning": { "label": "属于…", "note": "它是让策略安全变好的训练法。" }, "policy-gradient": { "label": "改进…", "note": "在梯度更新外加上步幅护栏。" }, "proximal-policy-optimization": { "label": "启发…", "note": "PPO 用更简单方式近似它。" }, "kullback-leibler-divergence": { "label": "用…限步", "note": "KL 衡量新旧策略差多远。" } } } } }, { "id": "tts", "name": "TTS", "layer": "L4", "era": "2010s", "publishedAt": "2026-06-06T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "voice-cloning" }, { "to": "multimodal" } ], "track": "today", "i18n": { "en": { "fullName": "Text-to-Speech", "factExplain": "Technology that turns written text into natural-sounding speech.", "humanExplain": "TTS is a tiny voice actor in your phone. Hand it a script, and it reads it out loud.\n\nIt speaks in voice assistants and GPS apps. It lets AI read out loud without a real human voice.", "humanExplainDisplay": "TTS is a ==tiny voice actor== in your phone.\nHand it a ==script==,\nand it reads it out loud.\n\nIt speaks in voice assistants and GPS apps.\nIt lets AI read out loud\nwithout a real human voice.", "relationsNarrative": "STT\nSTT is the reverse of TTS: it writes down speech instead of speaking text.\n\nVoice cloning\nVoice cloning gives TTS a chosen voice to speak with.\n\nMultimodal AI\nTTS is a Multimodal output. It lets AI express content with sound.", "relations": { "speech-to-text": { "label": "pairs with …", "note": "TTS turns text into voice. STT turns voice into text." }, "voice-cloning": { "label": "can use … for custom voices", "note": "Voice cloning sets the voice before TTS reads the text." }, "multimodal": { "label": "is a … output", "note": "TTS lets AI answer with sound, not just text." } } }, "zh": { "fullName": "文本转语音", "factExplain": "把文字内容合成为自然语音的技术。", "humanExplain": "你把台词塞进点歌机，它下一秒就替你开嗓；人不用出声，字自己就会“说话”。\n\n常用于语音助手和导航播报，让机器不必真人配音就能开口，也方便无障碍朗读。", "humanExplainDisplay": "你把台词塞进==点歌机==，\n它下一秒就替你==开嗓==；\n人不用出声，\n字自己就会“说话”。\n\n常用于语音助手和导航播报，\n让机器不必真人配音就能开口，\n也方便无障碍朗读。", "relationsNarrative": "Speech-to-text\n它和 STT 正好相反，一个负责说，一个负责听写。\n\nVoice cloning\n语音克隆能提供特定音色，让它按指定声音开口。\n\nMultimodal AI\n它是多模态输出的一种，让 AI 能用声音表达内容。", "relations": { "speech-to-text": { "label": "和…互为入口", "note": "一个把字变声，一个把声变字。" }, "voice-cloning": { "label": "可结合…定制声音", "note": "先定音色，再把文字念出来。" }, "multimodal": { "label": "属于…应用", "note": "它让模型输出不只停留在文字。" } } } } }, { "id": "turing-test", "name": "Turing-test", "layer": "L1", "era": "1950", "publishedAt": "2026-05-23T12:05:00Z", "relations": [ { "to": "llm" }, { "to": "agi" }, { "to": "reasoning-model" } ], "i18n": { "en": { "fullName": "Turing Test", "factExplain": "A test of whether people can tell a machine from a human in chat.", "humanExplain": "The Turing Test is AI’s fake-mustache interview. If you chat and cannot spot the robot, it passes.\n\nIt was a classic doorway into machine intelligence. Today, a model may chat like a person, but still not understand everything.", "humanExplainDisplay": "The Turing Test is AI’s ==fake-mustache interview==.\nIf you chat and cannot ==spot the robot==,\nit passes.\n\nIt was a classic doorway\ninto machine intelligence.\nToday, a model may chat like a person,\nbut still not understand everything.", "relationsNarrative": "LLM\nLLMs made machines sound more human in normal chat.\n\nAGI\nThe Turing Test is often used to discuss visible signs of AGI.\n\nReasoning-model\nReasoning-models improve machine reasoning in hard Q&A.", "relations": { "llm": { "label": "approached by …", "note": "LLMs make machine chat sound more human." }, "agi": { "label": "still far from …", "note": "Passing the Turing Test does not prove AGI." }, "reasoning-model": { "label": "strengthened by …", "note": "Reasoning-models help machines handle harder questions." } } }, "zh": { "fullName": "图灵测试", "factExplain": "用人类是否能分辨机器来判断机器智能表现的测试。", "humanExplain": "图灵测试像 AI 的“能不能装成人类”面试，重点不是它有没有灵魂，而是你会不会被骗过。\n\n它曾是智能讨论的经典入口，但今天的模型会聊天，不代表真的会理解一切。", "humanExplainDisplay": "图灵测试像 AI 的\n==能不能装成人类==\n面试。\n重点不是它有没有灵魂，\n而是你会不会被骗过。\n\n它曾是智能讨论的经典入口。\n但今天会聊天，\n不代表真的什么都懂。", "relationsNarrative": "LLM\nLLM 提升了机器在自然语言对话中的拟人表现。\n\nAGI\nTuring-test 常被用来讨论 AGI 是否具备可观察能力。\n\nReasoning-model\nReasoning-model 强化了机器在复杂问答中的推理表现。", "relations": { "llm": { "label": "已被…逼近" }, "agi": { "label": "距…仍远" }, "reasoning-model": { "label": "由…补强" } } } } }, { "id": "u-net", "name": "U-Net", "layer": "L3", "era": "2015", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "cnn" }, { "to": "computer-vision" }, { "to": "diffusion" }, { "to": "resnet" } ], "track": "history", "i18n": { "en": { "fullName": "U-Net Image Segmentation Network", "factExplain": "A CNN often used to split images into pixel-level parts.", "humanExplain": "U-Net is like a fussy kid with a coloring book. It colors only the cat, not the couch.\n\nIt is common in medical scan work. Diffusion models also use it to clean noise and rebuild images step by step.", "humanExplainDisplay": "U-Net is like a ==fussy kid==\nwith a coloring book.\nIt ==colors only the cat==,\nnot the couch.\n\nIt is common in medical scan work.\nDiffusion models also use it\nto clean noise and rebuild images\nstep by step.", "relationsNarrative": "CNN\nU-Net is a classic variant of a CNN.\n\nComputer Vision\nU-Net mainly helps with image segmentation in Computer Vision.\n\nDiffusion\nMany Diffusion models use U-Net as the main denoising network.\n\nResNet\nU-Net and ResNet are both classic vision designs, but they focus on different jobs.", "relations": { "cnn": { "label": "is a kind of …", "note": "U-Net is a classic variant of a CNN." }, "computer-vision": { "label": "serves … tasks", "note": "U-Net is often used for image segmentation in vision work." }, "diffusion": { "label": "often forms the backbone of …", "note": "Many diffusion models use U-Net as the denoising network." }, "resnet": { "label": "shares the vision family with …", "note": "Both are classic vision designs, but they focus on different jobs." } } }, "zh": { "fullName": "U-Net 图像分割网络", "factExplain": "一种常用于像素级图像分割的卷积神经网络。", "humanExplain": "它像医院拍片时拿荧光笔沿病灶一圈圈描边，连芝麻大的可疑阴影都要单独圈出来。\n\n常做医学影像分割，也常在扩散模型里负责逐步还原画面。", "humanExplainDisplay": "它像医院拍片时\n拿==荧光笔==沿病灶\n一圈圈==描边==，\n连芝麻大的可疑阴影\n都要单独圈出来。\n\n常做医学影像分割，\n也常在扩散模型里\n负责逐步还原画面。", "relationsNarrative": "CNN\n它是卷积神经网络的一种经典变体。\n\nComputer Vision\n它主要解决图像分割等计算机视觉问题。\n\nDiffusion\n很多扩散模型会用它充当去噪主干网络。\n\nResNet\n两者都属经典视觉架构，但侧重点不同。", "relations": { "cnn": { "label": "属于…一类", "note": "它本质上是卷积网络的变体。" }, "computer-vision": { "label": "服务…任务", "note": "它常用于图像分割等视觉场景。" }, "diffusion": { "label": "常作…骨架", "note": "很多扩散模型用它做去噪网络。" }, "resnet": { "label": "与…同属视觉架构", "note": "两者都是经典视觉模型设计。" } } } } }, { "id": "unification", "name": "Unification", "layer": "L2", "era": "1965", "publishedAt": "2026-06-18T04:00:00.000Z", "relations": [ { "to": "foundation-model" }, { "to": "multimodal" }, { "to": "agent" } ], "track": "history", "i18n": { "en": { "fullName": "Logical Unification", "factExplain": "A matching process that swaps variables so two logic expressions agree.", "humanExplain": "Unification is fill-in-the-blank homework for logic. Put “pizza” in the blank, and both answers line up.\n\nIt swaps a variable for a real item. This makes two logic sentences match in automatic reasoning.", "humanExplainDisplay": "Unification is ==fill-in-the-blank homework== for logic.\nPut “pizza” in the blank,\nand ==both answers line up==.\n\nIt swaps a variable for a real item.\nThis makes two logic sentences match\nin automatic reasoning.", "relationsNarrative": "Foundation-model\nUnification is symbol matching, not a shared model base.\n\nMultimodal AI\nMultimodal AI joins data types, but unification swaps variables in logic.\n\nAgent\nAn Agent can use unification to match rules with current facts.", "relations": { "foundation-model": { "label": "differs from …", "note": "Unification matches symbols in rules. It is not one shared model base." }, "multimodal": { "label": "differs from …", "note": "Multimodal AI joins data types. Unification matches variables in logic." }, "agent": { "label": "can help …", "note": "An Agent can use unification to match rules with facts." } } }, "zh": { "fullName": "逻辑合一", "factExplain": "让两个逻辑表达式通过变量替换变得一致的匹配过程。", "humanExplain": "合一像食堂打饭对口令：阿姨喊“来个荤菜”，你回“红烧肉”，这格就严丝合缝对上了。\n\n它把变量替成具体项，让两句逻辑说成一回事，常用于自动推理。", "humanExplainDisplay": "合一像食堂打饭对口令：\n阿姨喊“来个==荤菜==”，\n你回“==红烧肉==”，\n这格就严丝合缝对上了。\n\n它把变量替成具体项，\n让两句逻辑说成一回事，\n常用于自动推理。", "relationsNarrative": "Foundation-model\n逻辑合一是符号推理里的匹配机制，不等于基础模型的能力统一。\n\nMultimodal\n多模态强调跨文本、图像、语音协同，和合一不是同一层概念。\n\nAgent\nAgent 若接入规则系统做推理，可用合一来对齐规则中的变量与当前事实。", "relations": { "foundation-model": { "label": "不同于", "note": "逻辑合一是符号推理里的表达式匹配，不是把多任务收进同一模型底座。" }, "multimodal": { "label": "不同于", "note": "多模态是在统一处理多种数据类型，和逻辑合一的变量替换匹配不是一回事。" }, "agent": { "label": "可用于", "note": "代理若用符号推理做规划或问答时，可借助合一来匹配规则与事实。" } } } } }, { "id": "unitree-robotics", "name": "Unitree Robotics", "layer": "L5", "sublayer": "product", "era": "2016", "publishedAt": "2026-07-04T04:00:00.000Z", "relations": [ { "to": "robotics" }, { "to": "embodied-ai" }, { "to": "physical-ai-ai" }, { "to": "vision-language-action-model-vla" } ], "track": "history", "i18n": { "en": { "fullName": "Unitree Robotics", "factExplain": "A Chinese company that builds four-legged and humanoid robots.", "humanExplain": "Unitree is like the neighbor with a movie robot dog. It trots down the sidewalk, not just across YouTube.\n\nIt makes four-legged robots and humanoid robots. You meet them in inspections, labs, and classrooms.", "humanExplainDisplay": "Unitree is like the neighbor with a ==movie robot dog==.\nIt trots down the ==sidewalk==,\nnot just across YouTube.\n\nIt makes four-legged robots and humanoid robots.\nYou meet them in inspections, labs, and classrooms.", "relationsNarrative": "Robotics\nUnitree Robotics is a key hardware company in Robotics.\n\nEmbodied AI\nUnitree gives Embodied AI a real body that can run and jump.\n\nPhysical AI\nLower-cost robot dogs bring Physical AI closer to everyday people.\n\nVLA\nA VLA can help a robot turn vision and instructions into action.", "relations": { "robotics": { "label": "belongs in …", "note": "Unitree puts AI into machines that can move." }, "embodied-ai": { "label": "gives a body to …", "note": "Real robot bodies bring AI off the screen and onto the floor." }, "physical-ai-ai": { "label": "helps spread …", "note": "Lower-cost hardware makes Physical AI easier to try." }, "vision-language-action-model-vla": { "label": "can connect to …", "note": "A VLA can turn what the robot sees and hears into action." } } }, "zh": { "fullName": "宇树科技", "factExplain": "一家研发四足与人形机器人的中国公司。", "humanExplain": "宇树像把科幻机器狗牵到楼下遛弯：能跑能跳，不再隔屏看热闹。\n\n它做四足和人形机器人，服务巡检、科研和教育。", "humanExplainDisplay": "宇树像把科幻机器狗\n牵到==楼下遛弯==：\n能跑能跳，\n不再==隔屏看热闹==。\n\n它做四足和人形机器人，\n服务巡检、科研和教育。", "relationsNarrative": "Robotics\nUnitree Robotics 是机器人赛道的代表硬件公司。\n\nEmbodied AI\n它给具身智能提供能跑、能跳的真实身体。\n\nPhysical AI\n低成本机器狗让物理 AI 更接近大众。\n\nVLA\nVLA 可让机器人把视觉和指令转成动作。", "relations": { "robotics": { "label": "属于…赛道", "note": "它把 AI 能力装进可运动的机器。" }, "embodied-ai": { "label": "承载…落地", "note": "真实机身让智能从屏幕走到地面。" }, "physical-ai-ai": { "label": "推动…普及", "note": "低成本硬件降低物理 AI 门槛。" }, "vision-language-action-model-vla": { "label": "可接入…", "note": "VLA 可把看懂指令变成动作。" } } } } }, { "id": "unsupervised-learning", "name": "Unsupervised Learning", "layer": "L2", "era": "1950s", "publishedAt": "2026-06-08T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "self-supervised-learning" }, { "to": "representation-learning" }, { "to": "word2vec" } ], "track": "history", "i18n": { "en": { "fullName": "Unsupervised Learning", "factExplain": "A way for AI to find structure in data without human labels.", "humanExplain": "It is the new kid in the school cafeteria. No name tags, but it still finds the pudding-cup traders.\n\nIt finds patterns without labels. It groups similar data. It can shrink data too. It also helps pretraining.", "humanExplainDisplay": "It is the ==new kid== in the school cafeteria.\nNo name tags,\nbut it still finds the ==pudding-cup traders==.\n\nIt finds patterns without labels.\nIt groups similar data.\nIt can shrink data too.\nIt also helps pretraining.", "relationsNarrative": "Supervised Learning\nSupervised Learning uses labels; Unsupervised Learning finds structure without them.\n\nSSL\nSSL keeps the same idea, but it builds its own training targets.\n\nRepresentation Learning\nUnsupervised Learning often learns useful features before later tasks.\n\nWord2Vec\nWord2Vec is a classic early example of unsupervised-style representation learning.", "relations": { "supervised-learning": { "label": "often compared with …", "note": "One learns from labels. The other finds patterns first." }, "self-supervised-learning": { "label": "came before …", "note": "SSL is a newer branch of the same basic idea." }, "representation-learning": { "label": "sets up …", "note": "It often learns reusable data features first." }, "word2vec": { "label": "helped create methods like …", "note": "Word2Vec is a classic early result of this style." } } }, "zh": { "fullName": "无监督学习", "factExplain": "不依赖人工标签，从数据中学习结构的方法。", "humanExplain": "像搬进新小区看邻居，门上没贴标签，也能慢慢分出谁爱遛狗、谁总团购、谁天天深夜外卖。\n\n它不靠标签找规律，常用于聚类、降维和预训练。", "humanExplainDisplay": "像搬进新小区看邻居，\n门上没贴标签，\n也能慢慢分出谁爱==遛狗==、\n谁总团购、谁天天==深夜外卖==。\n\n它不靠标签找规律，\n常用于聚类、\n降维和预训练。", "relationsNarrative": "Supervised Learning\n它常拿来和监督学习对照：一个靠标签，一个先找结构。\n\nSelf-Supervised Learning\n自监督学习继承了它的思路，但多了人为构造训练目标。\n\nRepresentation Learning\n无监督学习常用于先学表征，给下游任务打底。\n\nWord2Vec\nWord2Vec 是无监督式表示学习的经典早期例子。", "relations": { "supervised-learning": { "label": "常与…对比", "note": "一个靠标签学，一个先自己找规律。" }, "self-supervised-learning": { "label": "是…的前身", "note": "自监督可看作无监督的现代延伸。" }, "representation-learning": { "label": "为…打基础", "note": "它常先学出可复用的数据表征。" }, "word2vec": { "label": "催生…这类方法", "note": "经典词向量就是代表性成果之一。" } } } } }, { "id": "upper-confidence-bound", "name": "UCB", "layer": "L2", "era": "1985", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "exploration-exploitation" }, { "to": "thompson-sampling" }, { "to": "reinforcement-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Upper Confidence Bound", "factExplain": "A picking rule that balances trying options with earning rewards under uncertainty.", "humanExplain": "UCB is like choosing lunch at school. You grab the safe pizza, but still risk the mystery taco once.\n\nIt picks the choice with the best hopeful score. It shows up in slot-machine-style tests and recommendation feeds.", "humanExplainDisplay": "UCB is like choosing lunch at school.\nYou grab the ==safe pizza==,\nbut still risk the ==mystery taco== once.\n\nIt picks the choice\nwith the best hopeful score.\nIt shows up in slot-machine-style tests\nand recommendation feeds.", "relationsNarrative": "Exploration-Exploitation Tradeoff\nUCB is a classic way to balance trying new choices and using known winners.\n\nThompson Sampling\nUCB and Thompson Sampling both handle the choice between testing and earning.\n\nRL\nUCB can be used in RL to choose the next action.", "relations": { "exploration-exploitation": { "label": "balances …", "note": "UCB is a classic way to manage trying versus earning." }, "thompson-sampling": { "label": "is compared with …", "note": "Both choose between safe rewards and useful tests." }, "reinforcement-learning": { "label": "chooses actions in …", "note": "UCB can guide which action an RL system tries next." } } }, "zh": { "fullName": "Upper Confidence Bound／置信上界算法", "factExplain": "在不确定性下平衡试错与收益的选臂策略。", "humanExplain": "UCB 像租房挑房：靠谱房接着看，消息少但卖相好的，也得抽空跑一趟，免得漏掉真宝藏。\n\n常用于多臂老虎机和推荐分发，让系统边试新项边控制损失。", "humanExplainDisplay": "UCB 像租房挑房：\n靠谱房接着看，\n消息少但卖相好的，\n也得抽空跑一趟，\n免得漏掉\n==真宝藏==。\n\n常用于多臂老虎机\n和推荐分发，\n让系统边试新项\n边控制损失。", "relationsNarrative": "Exploration-Exploitation Tradeoff\n它是平衡探索与利用的经典做法。\n\nThompson Sampling\n它和汤普森采样都在解决试错与收益的取舍。\n\nReinforcement Learning\n它可作为强化学习里选择动作的策略。", "relations": { "exploration-exploitation": { "label": "平衡…取舍", "note": "它是经典的探索利用策略。" }, "thompson-sampling": { "label": "常与…对比", "note": "两者都解决试错与收益平衡。" }, "reinforcement-learning": { "label": "用于…选动作", "note": "它常作强化学习中的选臂规则。" } } } } }, { "id": "user-identity-verification-for-ai", "name": "ID Verification", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "captcha" }, { "to": "ai-bot-traffic" }, { "to": "data-privacy" }, { "to": "agent-security" } ], "track": "today", "i18n": { "en": { "fullName": "AI Identity Verification", "factExplain": "A way for an AI service to check who a user really is.", "humanExplain": "AI identity verification is the bouncer at a school dance. No wristband, no sneaking in as “DefinitelyNotABot42.”\n\nYou meet it at login. You also meet it before API use or risky actions. It blocks abuse. It should touch as little private data as possible.", "humanExplainDisplay": "AI identity verification is the ==bouncer==\nat a school dance.\nNo wristband,\nno sneaking in as ==DefinitelyNotABot42==.\n\nYou meet it at login.\nYou also meet it before API use\nor risky actions.\nIt blocks abuse.\nIt should touch as little private data as possible.", "relationsNarrative": "CAPTCHA\nCAPTCHA is a common first gate for identity verification.\n\nAI bot traffic\nIt helps tell real users from automated spam.\n\nData-privacy\nVerifying identity can touch sensitive personal data.\n\nAgent Security\nThe more an agent can do, the more it must know who approved it.", "relations": { "captcha": { "label": "checks humans with …", "note": "CAPTCHA is often the first gate for identity checks." }, "ai-bot-traffic": { "label": "blocks …", "note": "Identity checks help cut down fake automated traffic." }, "data-privacy": { "label": "affects …", "note": "Deeper checks can raise privacy risk." }, "agent-security": { "label": "strengthens …", "note": "Agents need user checks before they get more power." } } }, "zh": { "fullName": "AI 用户身份验证", "factExplain": "在 AI 服务中确认用户真实身份的机制。", "humanExplain": "AI 身份验证像高铁检票：票证人对上，黄牛和机器小号才进不了站。\n\n用于登录、接口、高风险操作，防滥用同时少碰隐私。", "humanExplainDisplay": "AI 身份验证像高铁检票：\n==票证人对上==，\n黄牛和==机器小号==\n才进不了站。\n\n用于登录、接口、\n高风险操作，\n防滥用同时少碰隐私。", "relationsNarrative": "CAPTCHA\n验证码是用户身份验证最常见的入口门槛。\n\nAI Bot Traffic\n它帮助区分真人请求和自动化刷量。\n\nData Privacy\n验证用户身份常会触碰敏感个人信息。\n\nAgent Security\n代理越能办事，越需要确认是谁在授权。", "relations": { "captcha": { "label": "用…做人机校验", "note": "验证码常当第一道门槛。" }, "ai-bot-traffic": { "label": "拦截…", "note": "身份校验能压低机器流量。" }, "data-privacy": { "label": "牵动…", "note": "验得越细，隐私风险越高。" }, "agent-security": { "label": "加强…", "note": "确认用户后才敢放权给代理。" } } } } }, { "id": "value-iteration", "name": "Value Iteration", "layer": "L2", "era": "1957", "publishedAt": "2026-06-07T04:00:00.000Z", "relations": [ { "to": "bellman-equation" }, { "to": "dynamic-programming" }, { "to": "policy-iteration" }, { "to": "markov-decision-process" } ], "track": "history", "i18n": { "en": { "fullName": "Value Iteration", "factExplain": "A method that repeats value updates to find the best action plan.", "humanExplain": "Value Iteration is like putting score stickers on a maze. Each round, squares near the exit get better scores.\n\nIt keeps updating scores, then chooses the best next move. You meet it in planning and reinforcement learning.", "humanExplainDisplay": "Value Iteration is like putting ==score stickers== on a maze.\nEach round,\nsquares near the ==exit== get better scores.\n\nIt keeps updating scores,\nthen chooses the best next move.\nYou meet it in planning\nand reinforcement learning.", "relationsNarrative": "Bellman Equation\nValue Iteration repeats Bellman updates to reach the best values.\n\nDP\nValue Iteration is a classic DP method for step-by-step decisions.\n\nPolicy Iteration\nValue Iteration and Policy Iter. are both classic ways to find the best policy.\n\nMDP\nValue Iteration often solves an MDP when the rules are known.", "relations": { "bellman-equation": { "label": "repeats … updates", "note": "It uses Bellman updates to move closer to the best values." }, "dynamic-programming": { "label": "is a … method", "note": "Value Iteration is a classic DP algorithm." }, "policy-iteration": { "label": "sits beside …", "note": "Both methods find the best policy." }, "markov-decision-process": { "label": "solves …", "note": "It often solves MDPs when the rules are known." } } }, "zh": { "fullName": "价值迭代", "factExplain": "一种反复更新状态价值来求最优策略的方法。", "humanExplain": "像租房前把几套房的通勤、房租、采光反复盘：一轮轮算完，最值那套自然冒出来。\n\n常用于序贯决策中求最优策略，是规划和强化学习的经典方法。", "humanExplainDisplay": "像租房前把几套房的通勤、房租、采光\n反复==盘==：\n一轮轮算完，\n最值那套自然==冒出来==。\n\n常用于序贯决策中求最优策略，\n是规划和强化学习的经典方法。", "relationsNarrative": "Bellman-equation\n价值迭代通过反复做贝尔曼更新来逼近最优解。\n\nDynamic-programming\n它是动态规划在序贯决策问题中的经典做法。\n\nPolicy-iteration\n它和策略迭代并列，都是求最优策略的经典方法。\n\nMarkov-decision-process\n价值迭代常用于求解已知环境的马尔可夫决策过程。", "relations": { "bellman-equation": { "label": "反复应用…", "note": "它靠贝尔曼更新一步步逼近最优值。" }, "dynamic-programming": { "label": "属于…方法", "note": "价值迭代是动态规划的代表算法。" }, "policy-iteration": { "label": "和…并列", "note": "两者都用来求最优策略。" }, "markov-decision-process": { "label": "求解…", "note": "常用于求解已知转移的决策问题。" } } } } }, { "id": "variational-autoencoder", "name": "VAE", "layer": "L3", "era": "2013", "publishedAt": "2026-06-11T04:00:00.000Z", "relations": [ { "to": "generative-model" }, { "to": "embedding" }, { "to": "kullback-leibler-divergence" }, { "to": "gan" } ], "track": "history", "i18n": { "en": { "fullName": "Variational Autoencoder", "factExplain": "A neural network that learns a hidden pattern and makes new samples.", "humanExplain": "A VAE is like a smoothie shop with a tiny secret menu. It squeezes a big recipe into a flavor code, then blends a new drink from it.\n\nYou meet it in image work, data shrinking, and feature learning. It is steady, but fine details can look soft.", "humanExplainDisplay": "A VAE is like a smoothie shop\nwith a ==tiny secret menu==.\nIt squeezes a big recipe\ninto a ==flavor code==,\nthen blends a new drink from it.\n\nYou meet it in image work,\ndata shrinking,\nand feature learning.\nIt is steady,\nbut fine details can look soft.", "relationsNarrative": "Generative Model\nA VAE is a classic member of the generative model family.\n\nEmbedding\nIt compresses data into a hidden code you can sample.\n\nKL Divergence\nIt uses KL Divergence to keep the hidden space from getting wild.\n\nGAN\nVAEs are steadier, but GANs often make sharper images.", "relations": { "generative-model": { "label": "belongs to …", "note": "It is a classic generative model based on probability." }, "embedding": { "label": "learns an … space", "note": "Its middle code can work like an embedding." }, "kullback-leibler-divergence": { "label": "uses … to tame distribution", "note": "KL Divergence keeps the hidden space from getting too messy." }, "gan": { "label": "is often compared with …", "note": "VAEs are steadier, but GANs often make sharper images." } } }, "zh": { "fullName": "Variational Autoencoder（变分自编码器）", "factExplain": "一种学习潜在分布并生成样本的神经网络。", "humanExplain": "它像奶茶店把一长串配方先浓缩成底料，再按这个味型摇出一杯差不多但全新的。\n\n常用于图像生成、降维和表征学习，生成更稳但细节较软。", "humanExplainDisplay": "它像奶茶店把一长串配方\n先浓缩成==底料==，\n再按这个==味型==\n摇出一杯差不多\n但全新的。\n\n常用于图像生成、降维\n和表征学习，\n生成更稳但细节较软。", "relationsNarrative": "Generative Model\n它是生成模型家族里的经典成员。\n\nEmbedding\n它会把数据压到可采样的潜在表示里。\n\nKL Divergence\n它用 KL 散度约束潜在分布别太野。\n\nGan\n两者都能生成内容，但取舍不同。", "relations": { "generative-model": { "label": "属于…一类", "note": "它是经典概率生成模型之一。" }, "embedding": { "label": "学出…空间", "note": "中间那层潜在表示可当嵌入。" }, "kullback-leibler-divergence": { "label": "用…约束分布", "note": "训练时靠它把潜在空间拉规整。" }, "gan": { "label": "常与…对比", "note": "一个更稳，一个画面往往更锐。" } } } } }, { "id": "variational-inference", "name": "VI", "layer": "L2", "era": "1990s", "publishedAt": "2026-06-14T04:00:00.000Z", "relations": [ { "to": "kullback-leibler-divergence" }, { "to": "variational-autoencoder" }, { "to": "generative-model" } ], "track": "history", "i18n": { "en": { "fullName": "Variational Inference", "factExplain": "A way to replace hard probability math with a faster close-enough answer.", "humanExplain": "VI is like picking the fastest grocery line. You skip the coupon wizard and take the close-enough bet.\n\nIt helps train probability models and estimate uncertainty faster. The answer is useful, but not exact.", "humanExplainDisplay": "VI is like picking\n==the fastest grocery line==.\nYou skip ==the coupon wizard==\nand take the close-enough bet.\n\nIt helps train probability models\nand estimate uncertainty faster.\nThe answer is useful,\nbut not exact.", "relationsNarrative": "KL Divergence\nVariational Inference uses KL Divergence to measure the gap between its guess and the target.\n\nVAE\nVAE training uses Variational Inference as a core way to solve the hard part.\n\nGenerative Model\nIn a Generative Model, Variational Inference often guesses hidden variables after seeing data.", "relations": { "kullback-leibler-divergence": { "label": "measures gaps with …", "note": "KL Divergence shows how far the guess is from the true distribution." }, "variational-autoencoder": { "label": "helps train …", "note": "VAE training uses VI as a core shortcut for hard inference." }, "generative-model": { "label": "used for inference in …", "note": "Generative models use it to guess hidden variables after seeing data." } } }, "zh": { "fullName": "变分推断（Variational Inference）", "factExplain": "一种把复杂概率推断近似化的求解方法。", "humanExplain": "下棋碰上长考怪，没法把所有变化算穿，就先挑一条八成靠谱的主线走，总比卡死强。\n\n常用于概率模型训练和不确定性估计，求解更快，但答案是近似的。", "humanExplainDisplay": "下棋碰上长考怪，\n没法把所有变化\n算穿，\n就先挑一条\n==八成靠谱==的\n==主线走==；\n总比卡死强。\n\n常用于概率模型训练\n和不确定性估计，\n求解更快，\n但答案是近似的。", "relationsNarrative": "KL Divergence\n变分推断常用它衡量近似分布与目标分布的差距。\n\nVAE\nVAE 的训练目标里，变分推断是核心求解思路。\n\nGenerative Model\n在生成模型里，它常用于近似难算的隐藏变量后验。", "relations": { "kullback-leibler-divergence": { "label": "用…衡量差距", "note": "常靠它比较近似分布和真实分布。" }, "variational-autoencoder": { "label": "支撑…训练", "note": "它是这类生成模型的核心推断工具。" }, "generative-model": { "label": "用于…推断", "note": "常拿来近似隐藏变量的后验分布。" } } } } }, { "id": "vc-dimension", "name": "VC Dimension", "layer": "L2", "era": "1971", "publishedAt": "2026-06-17T04:00:00.000Z", "relations": [ { "to": "pac-learning" }, { "to": "statistical-learning-theory" }, { "to": "bias-variance-tradeoff" }, { "to": "regularization" } ], "track": "history", "i18n": { "en": { "fullName": "Vapnik–Chervonenkis Dimension", "factExplain": "A theory score for how complex a model can be.", "humanExplain": "VC Dimension is like testing a bendy lunchbox divider. It splits every weird pile of fries and peas. Powerful, but maybe too bendy.\n\nIt estimates how flexible a model is. It helps warn you about overfitting.", "humanExplainDisplay": "VC Dimension is like testing a ==bendy lunchbox divider==.\nIt splits every weird pile of ==fries and peas==.\nPowerful, but maybe too bendy.\n\nIt estimates how flexible a model is.\nIt helps warn you about overfitting.", "relationsNarrative": "PAC\nVC Dimension helps PAC set provable bounds on learning from data.\n\nSLT\nVC Dimension is a core measure of model complexity in SLT.\n\nBias-Variance Tradeoff\nA higher VC Dimension means more flexibility and more variance risk.\n\nRegularization\nRegularization limits effective complexity and helps reduce overfitting.", "relations": { "pac-learning": { "label": "supports … analysis", "note": "VC Dimension helps PAC prove limits on learning from samples." }, "statistical-learning-theory": { "label": "is core to …", "note": "VC Dimension is a key way SLT measures model complexity." }, "bias-variance-tradeoff": { "label": "explains …", "note": "A very flexible model often has high variance and overfits." }, "regularization": { "label": "helps explain …", "note": "Regularization helps keep model complexity under control." } } }, "zh": { "fullName": "VC 维度", "factExplain": "衡量模型能把数据分得多复杂的理论指标。", "humanExplain": "像裁缝看布料弹性：你怎么拉扯它都能兜住版型，说明料子真能打，也可能只是太容易跑偏变形。\n\n常用来估摸模型复杂度，判断它会不会太容易过拟合。", "humanExplainDisplay": "像裁缝看布料==弹性==：\n你怎么拉扯它\n都能兜住==版型==，\n说明料子真能打，也可能只是\n太容易跑偏变形。\n\n常用来估摸模型复杂度，\n判断它会不会\n太容易过拟合。", "relationsNarrative": "PAC Learning\n它常用来给 PAC 学习提供可证明的泛化界。\n\nStatistical Learning Theory\n它是统计学习理论里衡量假设空间复杂度的核心量。\n\nBias-Variance Tradeoff\nVC 维度越高，模型越灵活，也更容易出现高方差。\n\nRegularization\n正则化常被用来限制有效复杂度，缓解过拟合。", "relations": { "pac-learning": { "label": "支撑…分析", "note": "它是 PAC 学习里常见的复杂度刻画。" }, "statistical-learning-theory": { "label": "属于…核心", "note": "它是统计学习理论的重要概念。" }, "bias-variance-tradeoff": { "label": "解释…根源", "note": "模型太复杂时，更容易高方差过拟合。" }, "regularization": { "label": "帮助理解…", "note": "正则化常用于压住过高的模型复杂度。" } } } } }, { "id": "vector-db", "name": "Vector-db", "layer": "L5", "sublayer": "architecture", "era": "2023", "publishedAt": "2026-05-23T10:05:00Z", "relations": [ { "to": "embedding" }, { "to": "rag" }, { "to": "api" } ], "i18n": { "en": { "fullName": "Vector Database", "factExplain": "A database built to store and search vectors.", "humanExplain": "A vector database is like a lunchroom seating chart for ideas. Pizza sits near burgers, not near math homework.\n\nIt finds meaning matches fast for RAG. You also see it in recommendation feeds and smart search.", "humanExplainDisplay": "A vector database is like a ==lunchroom seating chart== for ideas.\n==Pizza sits near burgers==,\nnot near math homework.\n\nIt finds meaning matches fast for RAG.\nYou also see it in recommendation feeds\nand smart search.", "relationsNarrative": "Embedding\nA vector database stores and searches the vectors made by Embedding.\n\nRAG\nRAG uses a vector database to pull useful facts from a knowledge base.\n\nAPI\nAn API connects vector database search to business apps.", "relations": { "embedding": { "label": "stores … vectors", "note": "A vector database stores and searches vectors made by Embedding." }, "rag": { "label": "is the base for …", "note": "RAG uses a vector database to fetch the right source text." }, "api": { "label": "serves search through …", "note": "An API lets apps use vector database search." } } }, "zh": { "fullName": "向量数据库", "factExplain": "专门存储和检索向量数据的数据库。", "humanExplain": "向量数据库像快递驿站：不按姓名找包裹，按“长得像”把相似内容堆一起。\n\n它常给知识库、推荐和 RAG 当底座，让 AI 快速捞到相关资料。", "humanExplainDisplay": "向量数据库像==快递驿站==：\n不按姓名找包裹，\n按“==长得像==”把相似内容堆一起。\n\n它常给知识库、\n推荐和 RAG 当底座，\n让 AI 快速捞到相关资料。", "relationsNarrative": "Embedding\nVector-db 负责存储和检索 Embedding 生成的向量。\n\nRAG\nRAG 依赖 Vector-db 从知识库中取回相关资料。\n\nAPI\nAPI 可把 Vector-db 的检索能力接入业务系统。", "relations": { "embedding": { "label": "存储…" }, "rag": { "label": "是…的基础设施" }, "api": { "label": "通过…提供检索" } } } } }, { "id": "vector-search", "name": "Vector search", "layer": "L5", "sublayer": "architecture", "era": "2023", "publishedAt": "2026-05-29T16:08:01.212Z", "relations": [ { "to": "embedding" }, { "to": "vector-db" }, { "to": "rag" } ], "track": "today", "i18n": { "en": { "fullName": "Vector search", "factExplain": "A search method that finds content by meaning, not exact words.", "humanExplain": "Vector search is like a librarian for half-remembered thoughts. Say “the space cat thing,” and it still finds the right page.\n\nYou meet it in knowledge-base search and recommendations. RAG uses it to find nearby facts first.", "humanExplainDisplay": "Vector search is like a ==librarian==\nfor ==half-remembered thoughts==.\nSay “the space cat thing,”\nand it still finds the right page.\n\nYou meet it in knowledge-base search\nand recommendations.\nRAG uses it to find nearby facts first.", "relationsNarrative": "Embedding\nVector search uses Embedding to turn text or images into comparable vectors.\n\nVector-db\nVector-db is the usual storage and search base for vector search.\n\nRAG\nRAG often uses vector search to find material before the model writes an answer.", "relations": { "embedding": { "label": "uses … to show meaning", "note": "Embedding turns content into vectors, so distance can be compared." }, "vector-db": { "label": "searches inside …", "note": "A Vector-db stores vectors and finds close matches fast." }, "rag": { "label": "finds facts for …", "note": "RAG often uses vector search to pull in useful source material." } } }, "zh": { "fullName": "向量搜索", "factExplain": "按语义相似度查找内容的检索方法。", "humanExplain": "你忘了原句也没事，只要记得“差不多那个意思”，它就像懂你话外音的馆员，能把资料扒拉出来。\n\n常用于知识库搜索、推荐和 RAG，擅长找相近内容。", "humanExplainDisplay": "你忘了原句也没事，\n只要记得“差不多那个意思”，\n它就像懂你话外音的\n==馆员==，\n能把资料\n==扒拉出来==。\n\n常用于知识库搜索、\n推荐和 RAG，\n擅长找相近内容。", "relationsNarrative": "Embedding\nVector search 先依赖 Embedding 把文本或图片变成可比较的向量。\n\nVector-db\nVector-db 是 Vector search 常见的存储和检索底座。\n\nRAG\nRAG 往往先用 Vector search 找资料，再交给模型生成回答。", "relations": { "embedding": { "label": "靠…表示语义", "note": "先把内容变成向量，才能比远近。" }, "vector-db": { "label": "在…里查找", "note": "向量库负责存储索引并快速召回。" }, "rag": { "label": "为…找资料", "note": "RAG 常靠它先捞出相关材料。" } } } } }, { "id": "version-space-learning", "name": "Version Space Learning", "layer": "L2", "era": "1977", "publishedAt": "2026-07-07T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "inductive-bias" }, { "to": "active-learning" }, { "to": "pac-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Version Space Learning", "factExplain": "A set of candidate rules that still match the labeled examples.", "humanExplain": "Version Space Learning is like Guess Who? for AI models. Each labeled example flips down more little faces.\n\nYou meet it in supervised learning and active learning. It shows how examples shrink the model choices.", "humanExplainDisplay": "Version Space Learning is like ==Guess Who?==\nfor AI models.\nEach labeled example ==flips down==\nmore little faces.\n\nYou meet it in supervised learning\nand active learning.\nIt shows how examples shrink\nthe model choices.", "relationsNarrative": "Supervised Learning\nVersion Space Learning uses labeled examples to drop rules that do not fit.\n\nInductive Bias\nInductive Bias sets the starting pool of possible rules.\n\nActive Learning\nActive Learning can pick examples that shrink the version space fastest.\n\nPAC\nPAC also asks how many examples can rule out wrong rules.", "relations": { "supervised-learning": { "label": "belongs to …", "note": "Labels remove candidate rules that do not fit." }, "inductive-bias": { "label": "is limited by …", "note": "Bias decides which rules may enter the candidate set." }, "active-learning": { "label": "guides … sample choice", "note": "The best next example splits the candidates apart." }, "pac-learning": { "label": "connects to …", "note": "Both study how many examples remove wrong rules." } } }, "zh": { "fullName": "版本空间学习", "factExplain": "维护所有与样本标签一致的候选假设集合。", "humanExplain": "版本空间学习像相亲筛简历：条件不符先划掉，剩下全进候选池。\n\n用于监督和主动学习，观察样本如何缩小候选模型。", "humanExplainDisplay": "版本空间学习像相亲筛简历：\n==条件不符先划掉==，\n剩下全进==候选池==。\n\n用于监督和主动学习，\n观察样本如何，\n缩小候选模型。", "relationsNarrative": "Supervised Learning\n它用带标签样本不断筛掉不一致的假设。\n\nInductive Bias\n偏好和假设空间决定候选范围从哪开始。\n\nActive Learning\n主动学习常挑最能切分版本空间的样本。\n\nPAC\nPAC 也关心用多少样本排除错误假设。", "relations": { "supervised-learning": { "label": "属于…范式", "note": "它靠标签样本筛掉不合假设。" }, "inductive-bias": { "label": "受…限制", "note": "偏好决定哪些假设能进候选池。" }, "active-learning": { "label": "指导…选样本", "note": "分歧最大的样本最能缩小空间。" }, "pac-learning": { "label": "连接…理论", "note": "两者都研究样本如何约束假设。" } } } } }, { "id": "vibe-coding", "name": "Vibe-coding", "layer": "L6", "era": "2025", "publishedAt": "2026-05-23T10:50:00Z", "relations": [ { "to": "cursor" }, { "to": "copilot" }, { "to": "prompt" }, { "to": "agent" } ], "i18n": { "en": { "fullName": "Vibe-coding", "factExplain": "A way to build code by telling AI what you want in plain language.", "humanExplain": "Vibe-coding is like hiring a robot handyman for your app. You say, “Add a login. Make it shiny,” and it starts hammering.\n\nIt helps beginners make prototypes fast. It can also hand you bugs you cannot read.", "humanExplainDisplay": "Vibe-coding is like hiring a ==robot handyman==\nfor your app.\nYou say, “Add a login.\n==Make it shiny==,”\nand it starts hammering.\n\nIt helps beginners make prototypes fast.\nIt can also hand you bugs\nyou cannot read.", "relationsNarrative": "Cursor\nCursor lets Vibe-coding work inside real projects.\n\nCopilot\nCopilot lowers the bar for coding with plain language.\n\nPrompt\nA clear Prompt makes Vibe-coding match your goal better.\n\nAgent\nAn Agent can move Vibe-coding from code writing to task doing.", "relations": { "cursor": { "label": "depends on …", "note": "Cursor lets Vibe-coding edit real project files more directly." }, "copilot": { "label": "depends on …", "note": "Copilot makes coding with plain language easier to start." }, "prompt": { "label": "states needs with …", "note": "Clear prompts make Vibe-coding land closer to your goal." }, "agent": { "label": "expands through …", "note": "Agents can turn Vibe-coding from writing code into doing tasks." } } }, "zh": { "fullName": "氛围编程", "factExplain": "用自然语言指挥 AI 辅助完成代码开发的方式。", "humanExplain": "氛围编程像对着程序员许愿：这里做个登录，那里加点高级感，然后等 AI 开始施工。\n\n它能让非专业者更快做原型，也会让人更快遇到自己看不懂的 bug。", "humanExplainDisplay": "氛围编程像对着程序员\n==许愿==。\n这里做个登录，那里加点高级感，\n然后等 AI 施工。\n\n它能让非专业者更快做原型。\n也会让人更快遇到\n自己看不懂的 bug。", "relationsNarrative": "Cursor\nCursor 让 Vibe-coding 更容易直接作用于真实项目。\n\nCopilot\nCopilot 降低了用自然语言参与编程的门槛。\n\nPrompt\nPrompt 越具体，Vibe-coding 生成结果越接近需求。\n\nAgent\nAgent 让 Vibe-coding 从代码生成扩展到任务执行。", "relations": { "cursor": { "label": "依赖…" }, "copilot": { "label": "依赖…" }, "prompt": { "label": "用…表达需求" } } } } }, { "id": "vision-language-action-model-vla", "name": "VLA", "layer": "L3", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "embodied-ai" }, { "to": "computer-use" }, { "to": "world-model" } ], "track": "today", "i18n": { "en": { "fullName": "Vision-Language-Action model", "factExplain": "A model that combines seeing, understanding words, and controlling real actions.", "humanExplain": "VLA is a kitchen helper with eyes and hands. Say “pass the spatula,” and it reaches for the spatula, not the TV remote.\n\nRobots use it to see a room. They use it to understand a request. Then they move. You meet it in robots doing real-world jobs.", "humanExplainDisplay": "VLA is a ==kitchen helper==\nwith ==eyes and hands==.\nSay “pass the spatula,”\nand it reaches for the spatula,\nnot the TV remote.\n\nRobots use it to see a room.\nThey use it to understand a request.\nThen they move.\nYou meet it in robots doing real-world jobs.", "relationsNarrative": "Multimodal AI\nVLA adds action output to multimodal understanding.\n\nEmbodied AI\nVLA is a core model for many embodied systems.\n\nComputer use\nVLA extends computer use from screens into the real world.\n\nWorld model\nA world model can help VLA predict what its moves may cause.", "relations": { "multimodal": { "label": "adds action to …", "note": "It adds control output to multimodal understanding." }, "embodied-ai": { "label": "powers …", "note": "Many embodied systems use VLA to sense and act." }, "computer-use": { "label": "extends … into real life", "note": "It takes screen control toward physical-world control." }, "world-model": { "label": "plans with …", "note": "A world model can help predict what an action will cause." } } }, "zh": { "fullName": "Vision-Language-Action model｜视觉-语言-动作模型", "factExplain": "把视觉、语言理解和动作控制整合到一起的模型。", "humanExplain": "这不是只会听口令的机器人，是师傅一句“把扳手递来”，它真能看准再动手。\n\n常用于机器人干活，让机器看环境、懂指令，并执行现实动作。", "humanExplainDisplay": "这不是只会听口令的机器人，\n是师傅一句\n“把==扳手递来==”，\n它真能看准再==动手==。\n\n常用于机器人干活，\n让机器看环境、懂指令，\n并执行现实动作。", "relationsNarrative": "Multimodal\n它是在多模态理解基础上，再增加动作输出能力。\n\nEmbodied AI\nVLA 是很多具身智能系统里负责感知到行动的核心模型。\n\nComputer Use\n它把“会操作电脑”进一步扩展成“会操作现实世界”。\n\nWorld Model\n世界模型可帮助它预判动作后果，提升执行稳定性。", "relations": { "multimodal": { "label": "把…接上动作", "note": "它是在多模态基础上再连控制输出。" }, "embodied-ai": { "label": "作为…核心模型", "note": "很多具身系统靠它感知并行动。" }, "computer-use": { "label": "扩展…到现实", "note": "前者多在屏幕上，后者走向物理世界。" }, "world-model": { "label": "可配合…决策", "note": "世界模型帮它预测动作后的结果。" } } } } }, { "id": "vision-language-model-vlm", "name": "VLM", "layer": "L3", "era": "2020s", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "multimodal" }, { "to": "clip" }, { "to": "llm" }, { "to": "vision-language-action-model-vla" } ], "track": "today", "i18n": { "en": { "fullName": "Vision-Language Model", "factExplain": "A multimodal model that understands both images and text.", "humanExplain": "A VLM is the friend zooming in on your vacation photo. It reads the caption too, then spots the raccoon stealing chips.\n\nIt helps AI answer questions about pictures. It also powers image search and factory checks.", "humanExplainDisplay": "A VLM is the friend\n==zooming in== on your vacation photo.\nIt reads the caption too,\nthen spots the ==raccoon stealing chips==.\n\nIt helps AI answer questions about pictures.\nIt also powers image search\nand factory checks.", "relationsNarrative": "Multimodal AI\nA VLM is a common kind of Multimodal AI.\n\nCLIP\nCLIP helps match pictures with the right words.\n\nLLM\nA VLM often gives an LLM eyes, so it can answer about pictures.\n\nVLA\nVLA goes one step further and turns what it sees into actions.", "relations": { "multimodal": { "label": "is a kind of …", "note": "VLM is a common type of Multimodal AI." }, "clip": { "label": "builds on … alignment", "note": "CLIP made image-text matching a common path." }, "llm": { "label": "gives … eyes", "note": "A vision part lets the LLM answer questions about images." }, "vision-language-action-model-vla": { "label": "can grow into …", "note": "VLA adds action after understanding images and text." } } }, "zh": { "fullName": "视觉语言模型", "factExplain": "同时理解图像和文本的多模态模型。", "humanExplain": "VLM像相亲局里的细节控：看照片又读简介，连背景小狗都能聊出话题。\n\n用于看图问答、搜图和巡检，让机器先看懂再回应。", "humanExplainDisplay": "VLM像相亲局里的\n==细节控==：\n看照片又读简介，\n连==背景小狗==都能聊出话题。\n\n用于看图问答、\n搜图和巡检，\n让机器先看懂再回应。", "relationsNarrative": "Multimodal AI\n视觉语言模型是多模态 AI 的代表模型之一。\n\nCLIP\nCLIP 让图像和文字在同一语义空间对齐。\n\nLLM\n它常把视觉编码接到 LLM 上，回答图片问题。\n\nVLA\nVLA 在看懂图文后，进一步控制动作。", "relations": { "multimodal": { "label": "属于…", "note": "它是多模态 AI 的典型形态。" }, "clip": { "label": "继承…的图文对齐", "note": "CLIP 奠定图文对齐的常用路线。" }, "llm": { "label": "给…加上眼睛", "note": "接入视觉编码后，LLM 能看图回答。" }, "vision-language-action-model-vla": { "label": "扩展成…", "note": "VLA 在看懂图文后继续输出动作。" } } } } }, { "id": "vision-transformer", "name": "Vision Transformer", "layer": "L3", "era": "2020", "publishedAt": "2026-06-10T04:00:00.000Z", "relations": [ { "to": "transformer" }, { "to": "self-attention" }, { "to": "cnn" }, { "to": "clip" } ], "track": "history", "i18n": { "en": { "fullName": "Vision Transformer", "factExplain": "A vision model that cuts an image into patches and uses a Transformer.", "humanExplain": "A Vision Transformer treats a photo like a tray of brownies. It cuts it into squares, then judges the whole pan.\n\nIt is used to recognize and sort images. It also often acts as the eyes in models that handle images and text.", "humanExplainDisplay": "A Vision Transformer treats a photo\nlike a ==tray of brownies==.\nIt cuts it into squares,\nthen judges the ==whole pan==.\n\nIt is used to recognize\nand sort images.\nIt also often acts as the eyes\nin models that handle images and text.", "relationsNarrative": "Transformer\nA Vision Transformer brings the Transformer from text tasks into image tasks.\n\nSelf-Attention\nSelf-Attention lets it compare patches across the whole image at once.\n\nCNN\nA Vision Transformer leans global, while a CNN leans local.\n\nCLIP\nCLIP often uses it as the vision encoder for pictures.", "relations": { "transformer": { "label": "brings … to vision", "note": "It uses the Transformer design for images." }, "self-attention": { "label": "sees the whole image with …", "note": "Self-Attention helps it connect far-apart image patches." }, "cnn": { "label": "takes a different path from …", "note": "It looks more globally, while CNNs lean more local." }, "clip": { "label": "often acts as …'s vision tower", "note": "CLIP often uses it as the part that reads pictures." } } }, "zh": { "fullName": "视觉 Transformer", "factExplain": "把图像切成小块后用 Transformer 处理的视觉模型。", "humanExplain": "视觉Transformer像拼地图：先把整张图切成小块再一起看，不再只盯着眼前巴掌大一片。\n\n常用于看图分类识别，也常做多模态模型的视觉底座。", "humanExplainDisplay": "视觉Transformer像==拼地图==：\n先把整张图切成小块再一起看，\n不再只盯着==眼前巴掌大一片==。\n\n常用于看图分类识别，\n也常做多模态模型的视觉底座。", "relationsNarrative": "Transformer\n它把 Transformer 从文本世界搬到了图像任务里。\n\nSelf-Attention\nSelf-Attention 让它能同时关注整张图里的区域关系。\n\nCNN\n它和 CNN 路线不同：前者偏全局，后者偏局部。\n\nCLIP\nCLIP 常用它做视觉编码器，负责把图片变成表示。", "relations": { "transformer": { "label": "把…搬到视觉", "note": "它把语言里的这套结构用到图像上。" }, "self-attention": { "label": "靠…看全图", "note": "它用全局关注来建模图像区域关系。" }, "cnn": { "label": "和…路线不同", "note": "一个更看全局，一个更偏局部归纳。" }, "clip": { "label": "常作为…视觉塔", "note": "很多图文模型拿它来负责看图。" } } } } }, { "id": "voice-cloning", "name": "Voice cloning", "layer": "L4", "era": "2023", "publishedAt": "2026-05-31T00:57:33.086Z", "relations": [ { "to": "deepfake" }, { "to": "data-privacy" }, { "to": "ai-election-safeguards" } ], "track": "today", "i18n": { "en": { "fullName": "Voice cloning", "factExplain": "Technology that makes a voice copy from a few short recordings.", "humanExplain": "Give AI a few voice notes, and it can wear your voice like a Halloween mask. It says “Hi, Mom,” and Mom pauses mid-coffee.\n\nPeople use it for dubbing and accessibility. It can also help scammers fake calls and spread lies.", "humanExplainDisplay": "Give AI a few voice notes,\nand it can ==wear your voice==\nlike a ==Halloween mask==.\nIt says “Hi, Mom,”\nand Mom pauses mid-coffee.\n\nPeople use it for dubbing and accessibility.\nIt can also help scammers fake calls\nand spread lies.", "relationsNarrative": "Deepfake\nVoice cloning is a common Deepfake form for audio.\n\nData-privacy\nVoice cloning can involve private recordings and voiceprints.\n\nElection Guard\nRealistic fake voices can mislead people during elections.", "relations": { "deepfake": { "label": "is a kind of …", "note": "Fake voices are a common form of Deepfake." }, "data-privacy": { "label": "raises … risks", "note": "Voice samples can reveal your identity and voiceprint." }, "ai-election-safeguards": { "label": "tests …", "note": "Fake voices make election messages harder to trust." } } }, "zh": { "fullName": "声音克隆", "factExplain": "用少量录音生成近似本人声线的技术。", "humanExplain": "几句录音下肚，AI 就像借走了你的嗓子，张口那一下，亲妈都得愣两秒。\n\n常用于配音和无障碍内容；也会放大冒充诈骗与伪造传播风险。", "humanExplainDisplay": "几句录音下肚，\nAI 就像==借走了你的嗓子==，\n张口那一下，\n==亲妈都得愣两秒==。\n\n常用于配音和无障碍内容；\n也会放大冒充诈骗与伪造传播风险。", "relationsNarrative": "Deepfake\nVoice cloning 是 Deepfake 在声音上的典型表现形式。\n\nData-privacy\n训练或模仿他人声音，往往会牵涉声纹与录音隐私。\n\nAi-election-safeguards\n逼真的仿声内容，会提高选举场景中的误导风险。", "relations": { "deepfake": { "label": "属于…的一种", "note": "伪造声音是 Deepfake 的常见形式。" }, "data-privacy": { "label": "牵涉…风险", "note": "录音样本可能暴露个人身份与声纹。" }, "ai-election-safeguards": { "label": "挑战…机制", "note": "仿声内容会增加选举传播中的识别难度。" } } } } }, { "id": "voice-to-voice-ai", "name": "Voice-to-voice-ai", "layer": "L4", "era": "2020s", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "speech-to-text" }, { "to": "tts" }, { "to": "real-time-ai-translation" }, { "to": "multimodal" } ], "track": "today", "i18n": { "en": { "fullName": "Voice-to-voice AI", "factExplain": "An AI style that hears speech and answers with speech.", "humanExplain": "Voice-to-voice AI is like a super awake drive-thru clerk. You talk, and it answers before your fries get cold.\n\nYou meet it in voice assistants. Support lines use it too. So do speaking practice apps. The magic is fast back-and-forth talk.", "humanExplainDisplay": "Voice-to-voice AI is like a ==super awake drive-thru clerk==.\nYou talk,\nand it answers before your ==fries get cold==.\n\nYou meet it in voice assistants.\nSupport lines use it too.\nSo do speaking practice apps.\nThe magic is fast back-and-forth talk.", "relationsNarrative": "STT\nVoice-to-voice AI does not always need text as a middle step.\n\nTTS\nTTS gives it the spoken answer, so the chat feels complete.\n\nReal-time AI Translation\nReal-time AI Translation often uses it for live spoken chats.\n\nMultimodal AI\nVoice-to-voice AI is a kind of Multimodal AI built around sound.", "relations": { "speech-to-text": { "label": "skips the … relay", "note": "It may understand speech without turning it into text first." }, "tts": { "label": "closes the loop with …", "note": "Spoken output is the other half of the chat." }, "real-time-ai-translation": { "label": "powers … chats", "note": "Live translation often uses it as the talk layer." }, "multimodal": { "label": "is a form of …", "note": "Sound is its main input and output." } } }, "zh": { "fullName": "语音到语音 AI", "factExplain": "直接接收语音并输出语音的 AI 交互方式。", "humanExplain": "不是按一下收音机等半天，更像菜市场砍价老手当场接话：你刚开口，它立马回嘴，来回才像真聊天。\n\n常用于语音助手、客服和陪练，重点是实时接话、对话更自然。", "humanExplainDisplay": "不是按一下收音机\n等半天，\n更像==菜市场砍价老手==当场接话：\n你刚开口，它立马回嘴。\n\n常用于语音助手、客服\n和陪练，\n重点是实时接话、\n对话更自然。", "relationsNarrative": "Speech-to-text\n它不一定非要先转成文字，再继续理解和回答。\n\nTTS\n语音输出是它的另一半，没有这环就聊不起来。\n\nReal-time AI Translation\n跨语言实时对话，常把它当成交互外壳。\n\nMultimodal\n它属于多模态交互的一种，核心模态是声音。", "relations": { "speech-to-text": { "label": "绕开…中转", "note": "它不一定先转文字再理解。" }, "tts": { "label": "把…接成闭环", "note": "语音输出能力是体验关键一环。" }, "real-time-ai-translation": { "label": "支撑…交互", "note": "跨语种实时对话常靠它实现。" }, "multimodal": { "label": "属于…形态", "note": "它把声音作为核心输入输出。" } } } } }, { "id": "vq-vae", "name": "VQ-VAE", "layer": "L3", "era": "2017", "publishedAt": "2026-07-09T04:00:00.000Z", "relations": [ { "to": "variational-autoencoder" }, { "to": "autoencoder" }, { "to": "generative-model" }, { "to": "representation-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Vector-Quantized Variational Autoencoder", "factExplain": "A generative autoencoder that represents hidden data with discrete codes from a codebook.", "humanExplain": "VQ-VAE is like a school cafeteria using meal numbers. It saves “Meal 4,” not a fancy photo of your pizza.\n\nIt turns images or speech into small code numbers. Those codes help AI make new pictures or audio.", "humanExplainDisplay": "VQ-VAE is like a school cafeteria\nusing ==meal numbers==.\nIt saves “Meal 4,”\nnot a ==fancy photo== of your pizza.\n\nIt turns images or speech\ninto small code numbers.\nThose codes help AI make\nnew pictures or audio.", "relationsNarrative": "VAE\nVQ-VAE changes the VAE hidden space into a discrete codebook.\n\nAutoencoder\nIt still uses the compress first, rebuild later pattern.\n\nGenerative Model\nIts discrete codes can feed a later generative model.\n\nRepresentation Learning\nIt turns images and speech into reusable discrete codes.", "relations": { "variational-autoencoder": { "label": "reworks … hidden space", "note": "It changes the VAE hidden space into a discrete codebook." }, "autoencoder": { "label": "keeps … structure", "note": "It still compresses first, then rebuilds the input." }, "generative-model": { "label": "feeds …", "note": "Its codes give a generative model pieces to build from." }, "representation-learning": { "label": "learns … representations", "note": "It turns complex images or sound into reusable codes." } } }, "zh": { "fullName": "向量量化变分自编码器", "factExplain": "用离散码本表示潜变量的生成式自编码器。", "humanExplain": "VQ-VAE 像便利店扫码：不记每包零食细节，只存货号，回头按号补货复原。\n\n用于图像语音生成，把连续信号压成离散码。", "humanExplainDisplay": "VQ-VAE 像便利店扫码：\n不记每包零食细节，\n==只存货号==，\n回头==按号补货==复原。\n\n用于图像语音生成，\n把连续信号，\n压成离散码。", "relationsNarrative": "VAE\nVQ-VAE 把 VAE 的连续潜变量换成离散码本。\n\nAutoencoder\n它仍沿用先编码压缩、再解码重建的结构。\n\nGenerative Model\n它学到的离散编码常被后续生成模型使用。\n\nRepresentation Learning\n它把图像、语音压成更容易建模的离散表示。", "relations": { "variational-autoencoder": { "label": "改造…的潜变量", "note": "把 VAE 的潜空间改成离散码本。" }, "autoencoder": { "label": "继承…结构", "note": "仍是先压缩、再重建的自编码器。" }, "generative-model": { "label": "服务…建模", "note": "离散编码可作为生成模型的素材。" }, "representation-learning": { "label": "学习…表示", "note": "把复杂信号压成可复用的离散表示。" } } } } }, { "id": "vram", "name": "VRAM", "layer": "L5", "sublayer": "compute", "era": "2012", "publishedAt": "2026-05-29T16:08:01.210Z", "relations": [ { "to": "gpu" }, { "to": "quantization" }, { "to": "local-llm" }, { "to": "parameter" } ], "track": "today", "i18n": { "en": { "fullName": "Video RAM", "factExplain": "Fast memory a GPU uses to hold models and work data.", "humanExplain": "VRAM is the GPU’s desk. Big desk: the laptop and snacks fit. Tiny desk: one elbow starts a disaster.\n\nIt decides if a model runs smoothly on your computer. It also matters during training and when a model answers.", "humanExplainDisplay": "VRAM is the GPU’s ==desk==.\n==Big desk==: the laptop and snacks fit.\nTiny desk: one elbow starts a disaster.\n\nIt decides if a model runs smoothly on your computer.\nIt also matters during training\nand when a model answers.", "relationsNarrative": "GPU\nVRAM is the GPU’s closest memory, so many GPU limits are VRAM limits.\n\nQuantization\nQuantization shrinks a model, so it needs less VRAM.\n\nLocal-LLM\nWhen you run a Local-LLM, VRAM often decides how big it can be.\n\nParameter\nMore parameters usually use more VRAM during loading and work.", "relations": { "gpu": { "label": "serves as …'s desk", "note": "VRAM is the GPU’s local memory while it works." }, "quantization": { "label": "saves space with …", "note": "Quantization shrinks models, so they need less VRAM." }, "local-llm": { "label": "limits whether … runs locally", "note": "For a Local-LLM, VRAM often becomes the first wall." }, "parameter": { "label": "holds copies of …", "note": "More parameters usually need more VRAM." } } }, "zh": { "fullName": "显存", "factExplain": "GPU 用来存放模型和计算数据的高速内存。", "humanExplain": "显存这东西，像厨房台面：地方大，锅碗瓢盆都摆得开；地方小，再会做饭也得手忙脚乱腾地方。\n\n它决定模型能否顺畅运行，常见于本地部署、训练和推理。", "humanExplainDisplay": "显存这东西，像==厨房台面==：\n地方大，锅碗瓢盆都摆得开；\n地方小，再会做饭也得\n手忙脚乱==腾地方==。\n\n它决定模型能否顺畅运行，\n常见于本地部署、训练和推理。", "relationsNarrative": "GPU\n显存是 GPU 干活时最贴身的内存，很多算力限制本质上也是显存限制。\n\nQuantization\n量化会压缩模型所占空间，常被用来降低显存需求。\n\nLocal-LLM\n本地部署大模型时，显存大小常决定你能跑多大的模型。\n\nParameter\n模型参数越多，加载和计算时通常就越占显存。", "relations": { "gpu": { "label": "作为…工作台", "note": "显存是 GPU 执行计算时的本地内存。" }, "quantization": { "label": "靠…省空间", "note": "量化能减少显存占用，降低运行门槛。" }, "local-llm": { "label": "限制…能否本地跑", "note": "本地跑模型时，显存往往先成瓶颈。" }, "parameter": { "label": "装下…的副本", "note": "参数越多，通常越吃显存空间。" } } } } }, { "id": "wavenet", "name": "WaveNet", "layer": "L3", "era": "2016", "publishedAt": "2026-07-01T04:00:00.000Z", "relations": [ { "to": "tts" }, { "to": "autoregressive-model" }, { "to": "ai-audio-generation" }, { "to": "cnn" } ], "track": "history", "i18n": { "en": { "fullName": "WaveNet", "factExplain": "A neural network that creates raw sound waves one tiny point at a time.", "humanExplain": "WaveNet is like a patient kid drawing a sound wave with a crayon. One tiny wiggle, then the next.\n\nIt helped computer voices sound less robot-like. You meet its ideas in TTS and voice generation.", "humanExplainDisplay": "WaveNet is like a ==patient kid==\ndrawing a sound wave with a crayon.\n==One tiny wiggle==,\nthen the next.\n\nIt helped computer voices\nsound less robot-like.\nYou meet its ideas\nin TTS and voice generation.", "relationsNarrative": "TTS\nWaveNet helped TTS voices move from okay to more natural.\n\nAutoregressive Model\nWaveNet uses this idea to make audio one point after another.\n\nAudio Generation\nWaveNet was an important early model for neural audio generation.\n\nCNN\nWaveNet uses CNN-style layers to see more sound context.", "relations": { "tts": { "label": "makes … sound natural", "note": "WaveNet helped synthetic speech sound much more human." }, "autoregressive-model": { "label": "uses … ideas", "note": "It makes audio one point at a time, in order." }, "ai-audio-generation": { "label": "helped shape …", "note": "WaveNet was an early star in neural audio generation." }, "cnn": { "label": "borrows from …", "note": "CNN layers let WaveNet use a wider slice of sound." } } }, "zh": { "fullName": "波形生成网络", "factExplain": "一种逐点生成原始音频波形的神经网络模型。", "humanExplain": "WaveNet 像老师傅拉面：不端现成面，按顺序拉出声音波形。\n\n推动语音合成更自然，常见于配音和声音生成。", "humanExplainDisplay": "WaveNet 像\n==老师傅拉面==：\n不端现成面，\n按顺序拉出声音波形。\n\n推动语音合成更自然，\n常见于配音，\n和声音生成。", "relationsNarrative": "TTS\nWaveNet 曾把合成语音从“能听”推向更自然。\n\nAutoregressive Model\n它按时间顺序一点点生成下一段波形。\n\nAudio Generation\n它是神经音频生成的重要早期代表。\n\nCNN\n它用卷积结构扩大音频上下文感受野。", "relations": { "tts": { "label": "让…更自然", "note": "曾显著提升合成语音自然度。" }, "autoregressive-model": { "label": "采用…思路", "note": "按时间顺序逐点生成音频。" }, "ai-audio-generation": { "label": "推动…发展", "note": "它是早期神经音频生成代表。" }, "cnn": { "label": "借用…结构", "note": "用卷积扩大音频上下文感受野。" } } } } }, { "id": "webgpu", "name": "WebGPU", "layer": "L5", "sublayer": "compute", "era": "2023", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "gpu" }, { "to": "in-browser-ai-ai" }, { "to": "cuda" } ], "track": "today", "i18n": { "en": { "fullName": "WebGPU", "factExplain": "A standard that lets web pages use the GPU for graphics and general computing.", "humanExplain": "WebGPU is like giving a browser tab a forklift. Now it can move 3D furniture, not just hang posters.\n\nIt helps browser AI, image tools, and 3D run locally. That means less cloud work.", "humanExplainDisplay": "WebGPU is like giving a browser tab ==a forklift==.\nNow it can move ==3D furniture==,\nnot just hang posters.\n\nIt helps browser AI, image tools, and 3D run locally.\nThat means less cloud work.", "relationsNarrative": "GPU\nWebGPU lets web pages hand compute work to the local GPU.\n\nIn-browser AI\nWebGPU is a key base for speeding up models inside the browser.\n\nCUDA\nBoth work with GPUs, but WebGPU is built for the web and many platforms.", "relations": { "gpu": { "label": "uses … to speed up", "note": "WebGPU connects web page work to the local GPU." }, "in-browser-ai-ai": { "label": "helps power …", "note": "In-browser AI often uses WebGPU to run faster." }, "cuda": { "label": "compares with …", "note": "WebGPU is more cross-platform, but CUDA sits closer to the hardware." } } }, "zh": { "fullName": "网页 GPU 接口", "factExplain": "让网页调用 GPU 做图形与通用计算的标准接口。", "humanExplain": "WebGPU 像给网页接上工地塔吊：不只贴海报，还能搬砖搭 3D 楼。\n\n用于浏览器模型、图像处理和 3D，减少云端计算。", "humanExplainDisplay": "WebGPU 像给网页\n接上==工地塔吊==：\n不只贴海报，\n还能==搬砖搭 3D 楼==。\n\n用于浏览器模型、\n图像处理和 3D，\n减少云端计算。", "relationsNarrative": "GPU\nWebGPU 让网页把计算任务交给本机 GPU。\n\nIn-browser AI\n它是浏览器内模型加速的重要底座。\n\nCUDA\n两者都管 GPU，WebGPU 更偏网页和跨平台。", "relations": { "gpu": { "label": "调用…加速", "note": "它把网页计算接到本机 GPU 上。" }, "in-browser-ai-ai": { "label": "支撑…", "note": "浏览器内跑模型常靠它提速。" }, "cuda": { "label": "对标…", "note": "它更跨平台，但不如 CUDA 贴近硬件。" } } } } }, { "id": "wechat-xiaowei", "name": "Xiaowei", "layer": "L5", "sublayer": "product", "era": "2025", "publishedAt": "2026-06-22T04:00:00.000Z", "relations": [ { "to": "ai-super-app" }, { "to": "personal-ai-apps" }, { "to": "llm" } ], "track": "today", "i18n": { "en": { "fullName": "WeChat AI assistant “Xiaowei”", "factExplain": "A built-in chat AI assistant inside WeChat.", "humanExplain": "Xiaowei is the friend in your WeChat group with a clipboard. Ask once, and it looks things up or nudges tasks along.\n\nIt brings search and answers into WeChat. It also connects simple services, so AI feels easy to use.", "humanExplainDisplay": "Xiaowei is the ==friend in your WeChat group==\nwith a ==clipboard==.\nAsk once,\nand it looks things up\nor nudges tasks along.\n\nIt brings search and answers into WeChat.\nIt also connects simple services,\nso AI feels easy to use.", "relationsNarrative": "AI super app\nXiaowei makes WeChat feel more like an AI super app.\n\nPersonal AI apps\nXiaowei is a personal AI app for daily questions and small tasks.\n\nLLM\nXiaowei uses an LLM to understand you and write replies.", "relations": { "ai-super-app": { "label": "pushes WeChat toward …", "note": "Xiaowei puts AI inside an app people already use every day." }, "personal-ai-apps": { "label": "belongs to …", "note": "It helps people ask questions and handle daily tasks." }, "llm": { "label": "runs on …", "note": "The LLM understands the request and writes the reply." } } }, "zh": { "fullName": "微信 AI 助手“小微”", "factExplain": "微信内置的对话式 AI 助手。", "humanExplain": "小微像微信里的班主任课代表：你喊一声，查资料、发通知、跑腿都接茬。\n\n把搜索、问答、服务接进微信，让普通人更容易用 AI。", "humanExplainDisplay": "小微像微信里的\n==班主任课代表==：\n你喊一声，\n查资料、发通知、跑腿都接茬。\n\n把搜索、问答、\n服务接进微信，\n让普通人更容易用 AI。", "relationsNarrative": "AI Super App\n小微让微信具备更明显的 AI 超级入口属性。\n\nPersonal AI Apps\n它是面向个人日常问答和办事的 AI 应用。\n\nLLM\n它依赖大模型理解输入，并生成自然语言回复。", "relations": { "ai-super-app": { "label": "强化…形态", "note": "微信借助手把 AI 放进高频入口。" }, "personal-ai-apps": { "label": "属于…", "note": "它面向个人日常问答和办事。" }, "llm": { "label": "由…驱动", "note": "大模型负责理解输入并生成回复。" } } } } }, { "id": "weight-decay", "name": "Weight Decay", "layer": "L2", "era": "1987", "publishedAt": "2026-06-15T04:00:00.000Z", "relations": [ { "to": "regularization" }, { "to": "parameter" }, { "to": "adam" }, { "to": "lasso" } ], "track": "history", "i18n": { "en": { "fullName": "Weight Decay", "factExplain": "A regularization method that shrinks parameters to reduce overfitting.", "humanExplain": "Weight decay is a dog trainer for AI knobs. It says, “Easy, buddy. No need to bark at every leaf.”\n\nDuring training, it gently pushes big weights down. This helps the model learn patterns, not memorize every weird example.", "humanExplainDisplay": "Weight decay is a ==dog trainer==\nfor AI knobs.\nIt says, “Easy, buddy.\nNo need to ==bark at every leaf==.”\n\nDuring training,\nit gently pushes big weights down.\nThis helps the model learn patterns,\nnot memorize every weird example.", "relationsNarrative": "Regularization\nWeight decay is a form of regularization used to control overfitting.\n\nParameter\nWeight decay acts on parameters and nudges large weights smaller.\n\nAdam\nWeight decay is often set with Adam during training.\n\nLasso\nBoth limit parameters, but Lasso tends to make many weights zero.", "relations": { "regularization": { "label": "is a kind of …", "note": "It is a classic way to control overfitting." }, "parameter": { "label": "shrinks …", "note": "It works by punishing parameters that get too large." }, "adam": { "label": "is often set with …", "note": "Modern training often sets weight decay beside the optimizer." }, "lasso": { "label": "is often compared with …", "note": "Both limit parameters, but they push them in different ways." } } }, "zh": { "fullName": "权重衰减", "factExplain": "一种通过压小参数来抑制过拟合的正则化方法。", "humanExplain": "权重衰减像老中医下药：猛药别一把一把抓，剂量收着点，身体才不至于见啥症状都乱反应。\n\n常随训练一起用，帮模型少硬记、多学规律，降低过拟合。", "humanExplainDisplay": "权重衰减像老中医下药：\n猛药别一把一把抓，\n==剂量收着点==，\n身体才不至于\n见啥症状都==乱反应==。\n\n常随训练一起用，\n帮模型少硬记、多学规律，\n降低过拟合。", "relationsNarrative": "Regularization\n它是正则化的一种具体做法，用来控制过拟合。\n\nParameter\n它直接作用在参数上，倾向把权重压小。\n\nAdam\n训练时常和 Adam 一起配置，但实现细节会不同。\n\nLasso\n两者都限制参数大小，但 Lasso 更偏向压出稀疏性。", "relations": { "regularization": { "label": "属于…手段", "note": "它是控制过拟合的经典办法。" }, "parameter": { "label": "直接约束…", "note": "它通过惩罚过大的参数生效。" }, "adam": { "label": "常配合…使用", "note": "现代训练里常随优化器一起设定。" }, "lasso": { "label": "常被拿来对比", "note": "两者都限参数，但方式不同。" } } } } }, { "id": "weight-initialization", "name": "Weight Initialization", "layer": "L2", "era": "1980s", "publishedAt": "2026-06-28T04:00:00.000Z", "relations": [ { "to": "parameter" }, { "to": "gradient-descent" }, { "to": "backpropagation" }, { "to": "neural-network" } ], "track": "history", "i18n": { "en": { "fullName": "Weight Initialization", "factExplain": "A way to set model parameters before training starts.", "humanExplain": "Weight initialization is like starting a bike in the right gear. Too high means wobbling. Too low means hamster pedaling.\n\nIt sets the model’s first numbers. It affects training speed and stability in deep networks.", "humanExplainDisplay": "Weight initialization is like starting a bike\nin the ==right gear==.\nToo high means wobbling.\nToo low means ==hamster pedaling==.\n\nIt sets the model’s first numbers.\nIt affects training speed and stability\nin deep networks.", "relationsNarrative": "Parameter\nWeight initialization sets starting values for parameters before training.\n\nGradient Descent\nWeight initialization gives Gradient Descent its starting point.\n\nBackpropagation\nGood starting values help Backpropagation keep gradients stable.\n\nNeural-network\nA neural network needs weight initialization before training starts.", "relations": { "parameter": { "label": "sets starting values for …", "note": "It gives each parameter a starting value before training." }, "gradient-descent": { "label": "gives … a starting point", "note": "The starting values decide where Gradient Descent begins." }, "backpropagation": { "label": "affects … stability", "note": "Bad starting values can make gradients too small or too large." }, "neural-network": { "label": "prepares … for training", "note": "A neural network needs initial weights before training starts." } } }, "zh": { "fullName": "权重初始化", "factExplain": "为模型参数设置训练前初始数值的方法。", "humanExplain": "权重初始化像和面先兑水：太干太稀都难救，开局对了才好揉。\n\n影响训练速度和稳定性，常见于深度网络从零训练。", "humanExplainDisplay": "权重初始化像\n和面先兑水：\n==太干太稀都难救==，\n开局对了才好揉。\n\n影响训练速度和稳定性，\n常见于深度网络\n从零训练。", "relationsNarrative": "Parameter\n权重初始化在训练前为参数设定起始数值。\n\nGradient Descent\n它给梯度下降一个出发点，影响收敛路径。\n\nBackpropagation\n合适初值能让反向传播的梯度更稳定。\n\nNeural Network\n神经网络训练前需要先初始化权重。", "relations": { "parameter": { "label": "设置…初值", "note": "训练开始前先给参数一个起点。" }, "gradient-descent": { "label": "给…出发点", "note": "初值决定优化从哪里出发。" }, "backpropagation": { "label": "影响…稳定性", "note": "不当初值会让梯度过小或过大。" }, "neural-network": { "label": "服务…训练", "note": "神经网络训练前需要先初始化权重。" } } } } }, { "id": "widrow-hoff-learning-rule", "name": "Widrow-Hoff Learning Rule (Least Mean Squares Rule)", "layer": "L2", "era": "1960", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "supervised-learning" }, { "to": "gradient-descent" }, { "to": "perceptron" }, { "to": "sgd" } ], "track": "history", "i18n": { "en": { "fullName": "Widrow-Hoff Learning Rule (Least Mean Squares Rule)", "factExplain": "A supervised rule for updating weights to shrink squared errors.", "humanExplain": "It is like a bowling coach with a tiny steering wheel. A small miss gets a small turn. A huge miss gets more turning.\n\nIt fixes weights after each mistake in early neural nets and adaptive filters. Its tiny steps helped inspire Gradient Descent training.", "humanExplainDisplay": "It is like a ==bowling coach==\nwith a tiny steering wheel.\nA small miss gets a small turn.\nA huge miss gets ==more turning==.\n\nIt fixes weights after each mistake\nin early neural nets and adaptive filters.\nIts tiny steps helped inspire\nGradient Descent training.", "relationsNarrative": "Supervised Learning\nThe rule uses the right answer to find the error, then updates weights.\n\nGradient Descent\nThe rule moves weights downhill on squared error.\n\nPerceptron\nThe rule helped early linear neurons learn from each mistake.\n\nSGD\nIts one-example-at-a-time tiny updates pointed toward later SGD.", "relations": { "supervised-learning": { "label": "belongs to …", "note": "It uses the right answer to create an error signal." }, "gradient-descent": { "label": "updates weights with …", "note": "It nudges weights in a direction that lowers squared error." }, "perceptron": { "label": "improves early … training", "note": "It lets a linear neuron learn from how wrong it was." }, "sgd": { "label": "helped inspire …", "note": "Its one-example-at-a-time updates became a common idea." } } }, "zh": { "fullName": "Widrow-Hoff 学习规则（最小均方规则）", "factExplain": "一种按误差平方最小化更新权重的监督学习规则。", "humanExplain": "它像数学老师批卷：不只画叉，还按差几分把学生一点点掰回来。\n\n用于早期神经网络和滤波，启发梯度下降训练。", "humanExplainDisplay": "它像数学老师批卷：\n==不只画叉==，\n还==按差几分==，\n一点点掰回来。\n\n用于早期神经网络和滤波，\n启发梯度下降训练。", "relationsNarrative": "Supervised Learning\n它依赖标准答案计算误差，再据此更新权重。\n\nGradient Descent\n它本质上沿误差平方的下降方向调整参数。\n\nPerceptron\n它推动了早期线性神经元从犯错中学习。\n\nSGD\n它的逐样本小步更新，预示了后来的 SGD。", "relations": { "supervised-learning": { "label": "属于…方法", "note": "它用标准答案产生误差信号。" }, "gradient-descent": { "label": "用…更新权重", "note": "它按误差梯度一步步调整。" }, "perceptron": { "label": "改进早期…训练", "note": "它让线性神经元按误差大小学习。" }, "sgd": { "label": "启发…思想", "note": "逐样本小步更新后来很常见。" } } } } }, { "id": "wmt-shared-task", "name": "WMT Shared Task", "layer": "L4", "era": "2006", "publishedAt": "2026-07-08T04:00:00.000Z", "relations": [ { "to": "machine-translation" }, { "to": "bleu" }, { "to": "neural-machine-translation" }, { "to": "model-leaderboard" } ], "track": "history", "i18n": { "en": { "fullName": "Workshop on Machine Translation Shared Task", "factExplain": "A yearly test comparing machine translation systems on the same data.", "humanExplain": "WMT is like a spelling bee for translation bots. Everyone gets the same sentences, so nobody can hide behind the dictionary.\n\nResearchers use it to test MT systems. The results show the ranking and when new methods took over.", "humanExplainDisplay": "WMT is like a ==spelling bee==\nfor translation bots.\nEveryone gets the ==same sentences==,\nso nobody can hide\nbehind the dictionary.\n\nResearchers use it\nto test MT systems.\nThe results show the ranking\nand when new methods took over.", "relationsNarrative": "MT\nWMT gives MT systems the same test.\n\nBLEU\nBLEU was once a common automatic score in WMT.\n\nNMT\nNMT moved ahead of older methods in WMT over time.\n\nLeaderboard\nWMT results often become a Leaderboard for systems.", "relations": { "machine-translation": { "label": "tests … systems", "note": "WMT compares MT systems on the same test." }, "bleu": { "label": "scores with …", "note": "BLEU was once a common automatic score in WMT." }, "neural-machine-translation": { "label": "tracked the rise of …", "note": "WMT showed NMT replacing older translation methods fast." }, "model-leaderboard": { "label": "creates a …", "note": "Shared task results often become a ranked list of systems." } } }, "zh": { "fullName": "机器翻译共享评测任务", "factExplain": "机器翻译系统在统一数据上的年度评测任务。", "humanExplain": "WMT 像翻译机器人的统考：同一套题大家一起做，谁翻得准当场见分晓。\n\n用于机器翻译评测，记录方法换代和排名。", "humanExplainDisplay": "WMT 像翻译机器人的\n==统考==：\n同一套题大家一起做，\n谁翻得准==当场见分晓==。\n\n用于机器翻译评测，\n记录方法换代和排名。", "relationsNarrative": "Machine Translation\nWMT Shared Task 为机器翻译系统提供统一赛题。\n\nBLEU\nBLEU 曾是共享任务里常用的自动评分指标。\n\nNeural Machine Translation\n神经机器翻译在 WMT 中逐步超越旧方法。\n\nLeaderboard\n共享任务结果常被整理成系统排行榜。", "relations": { "machine-translation": { "label": "评测…系统", "note": "同题比较不同机器翻译系统。" }, "bleu": { "label": "常用…打分", "note": "BLEU 曾是机器翻译常用自动指标。" }, "neural-machine-translation": { "label": "见证…崛起", "note": "WMT 记录了神经翻译的快速替代。" }, "model-leaderboard": { "label": "形成…排名", "note": "共享任务常把系统结果排成榜单。" } } } } }, { "id": "word-error-rate", "name": "WER", "layer": "L6", "era": "1980s", "publishedAt": "2026-06-26T04:00:00.000Z", "relations": [ { "to": "speech-recognition" }, { "to": "speech-to-text" }, { "to": "connectionist-temporal-classification" } ], "track": "history", "i18n": { "en": { "fullName": "Word Error Rate", "factExplain": "A score that shows how much a transcript differs from the correct words.", "humanExplain": "WER is like a strict spelling-test teacher with one red pen. Every wrong, missing, or random extra word costs points.\n\nIt tests speech-to-text and captions. A lower WER means the transcript is closer to the real words.", "humanExplainDisplay": "WER is like a ==strict spelling-test teacher==\nwith one red pen.\nEvery ==wrong, missing, or random extra word==\ncosts points.\n\nIt tests speech-to-text and captions.\nA lower WER means the transcript\nis closer to the real words.", "relationsNarrative": "ASR\nASR uses WER to show how many words the transcript got wrong.\n\nSTT\nWhen STT matches the real words better, WER goes down.\n\nCTC\nCTC often trains speech models, and WER can check their output.", "relations": { "speech-recognition": { "label": "checks … accuracy", "note": "ASR uses WER to count transcript errors." }, "speech-to-text": { "label": "measures … transcript quality", "note": "The closer STT is to the real words, the lower WER is." }, "connectionist-temporal-classification": { "label": "tests … output", "note": "Speech models trained with CTC are often checked with WER." } } }, "zh": { "fullName": "词错误率", "factExplain": "衡量转写文本与标准答案差异的指标。", "humanExplain": "WER像语文老师改听写：错字、漏字、多写字，全记进同一本扣分账。\n\n常评估语音识别和字幕转写，数值越低越准。", "humanExplainDisplay": "WER像语文老师改听写：\n==错字、漏字、多写字==，\n全记进同一本==扣分账==。\n\n常评估语音识别和字幕转写，\n数值越低越准。", "relationsNarrative": "Speech Recognition\n语音识别常用 WER 衡量转写文本错了多少。\n\nSpeech-To-Text\nSTT 的转写结果越接近原文，WER 通常越低。\n\nConnectionist Temporal Classification\nCTC 常用于语音模型训练，输出可用 WER 检验。", "relations": { "speech-recognition": { "label": "评估…准确率", "note": "语音识别常用它衡量转写错误。" }, "speech-to-text": { "label": "衡量…转写质量", "note": "转文字结果越接近原文，数值越低。" }, "connectionist-temporal-classification": { "label": "检验…输出", "note": "CTC 训练的语音模型常用它评测。" } } } } }, { "id": "word-sense-disambiguation", "name": "WSD", "layer": "L4", "era": "1950", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "natural-language-understanding" }, { "to": "wordnet" }, { "to": "bert" }, { "to": "machine-translation" } ], "track": "history", "i18n": { "en": { "fullName": "Word Sense Disambiguation", "factExplain": "An NLP task for picking a word’s meaning from nearby words.", "humanExplain": "WSD is a nosy friend. You say “bank,” and it looks for a fishing rod or a debit card.\n\nYou meet it in search and translation. It keeps Q&A bots from chasing the wrong meaning.", "humanExplainDisplay": "WSD is a ==nosy friend==.\nYou say “bank,”\nand it looks for a ==fishing rod or a debit card==.\n\nYou meet it in search and translation.\nIt keeps Q&A bots from chasing the wrong meaning.", "relationsNarrative": "NLU\nWSD clears up word meaning before a sentence can be understood.\n\nWordNet\nWordNet often gives WSD the list of possible meanings.\n\nBERT\nBERT helps WSD use nearby words to pick the right meaning.\n\nMT\nWSD helps MT choose the right word in another language.", "relations": { "natural-language-understanding": { "label": "supports …", "note": "Words need clear meanings before sentences make sense." }, "wordnet": { "label": "uses … for meanings", "note": "WordNet often acts as the menu of possible meanings." }, "bert": { "label": "reads context with …", "note": "BERT helps WSD use context more accurately." }, "machine-translation": { "label": "helps … choose words", "note": "If WSD picks wrong, the translation goes off track." } } }, "zh": { "fullName": "词义消歧", "factExplain": "判断词语在具体语境中含义的NLP任务。", "humanExplain": "词义消歧像老中医号脉看“火”：是上火、火锅，还是火了，全靠上下文。\n\n用于搜索、翻译和问答，避免机器把词义带偏。", "humanExplainDisplay": "词义消歧像老中医\n号脉看==“火”==：\n是上火、火锅，还是火了，\n全靠上下文。\n\n用于搜索、翻译和问答，\n避免机器\n把词义带偏。", "relationsNarrative": "Natural Language Understanding\n词义消歧是理解句意前的基础小关卡。\n\nWordNet\nWordNet常提供可选择的词义集合。\n\nBERT\nBERT用上下文表示帮助判断当前词义。\n\nMachine Translation\n翻译前判清词义，才能选对目标词。", "relations": { "natural-language-understanding": { "label": "支撑…", "note": "先懂词义，才谈理解句子。" }, "wordnet": { "label": "借…定义词义", "note": "WordNet常充当词义清单。" }, "bert": { "label": "用…看上下文", "note": "上下文表示让消歧更准。" }, "machine-translation": { "label": "帮助…选词", "note": "词义判错，翻译就跑偏。" } } } } }, { "id": "word2vec", "name": "Word2Vec", "layer": "L3", "era": "2013", "publishedAt": "2026-06-04T04:00:00.000Z", "relations": [ { "to": "embedding" }, { "to": "vector-search" }, { "to": "neural-network" }, { "to": "bert" } ], "track": "history", "i18n": { "en": { "fullName": "Word2Vec", "factExplain": "A classic method for learning words as vectors.", "humanExplain": "Word2Vec is like watching the school cafeteria. Kids who sit together every day probably belong to the same crew.\n\nIt turns words into vectors, so computers can compare their distance. It powered semantic search. It also helped recommendations and early text tools.", "humanExplainDisplay": "Word2Vec is like watching the ==school cafeteria==.\nKids who ==sit together every day==\nprobably belong to the same crew.\n\nIt turns words into vectors,\nso computers can compare their distance.\nIt powered semantic search.\nIt also helped recommendations\nand early text tools.", "relationsNarrative": "Embedding\nWord2Vec is a classic early form of Embedding.\n\nVector search\nWord2Vec turns words into vectors, so Vector search can find similar meanings.\n\nNeural-network\nWord2Vec uses a shallow Neural-network to learn from words seen together.\n\nBERT\nBERT upgrades fixed word vectors by reading the surrounding words.", "relations": { "embedding": { "label": "is an early form of …", "note": "Word2Vec is a classic word embedding method." }, "vector-search": { "label": "gives vectors to …", "note": "Words need vectors before search can compare meanings." }, "neural-network": { "label": "learns vectors with …", "note": "Word2Vec uses a shallow neural network to learn word vectors." }, "bert": { "label": "was upgraded by …", "note": "BERT lets word meaning change with context." } } }, "zh": { "fullName": "词向量模型", "factExplain": "一种把词学习成向量表示的经典方法。", "humanExplain": "它像夜市里常拼桌的食客：老一起出现的，摊主一眼就知道他们八成是一路人，座位也越排越近。\n\n它把词变成可比较距离的向量，常用于语义搜索、推荐和早期文本表示。", "humanExplainDisplay": "它像夜市里常拼桌的食客：\n老一起出现的，\n摊主一眼就知道他们\n八成是==一路人==，\n座位也==越排越近==。\n\n它把词变成可比较距离的向量，\n常用于语义搜索、\n推荐和早期文本表示。", "relationsNarrative": "Embedding\n它是 Embedding 的经典早期形式，把词变成向量。\n\nVector search\n词先变成向量，才能做相似度检索和语义搜索。\n\nNeural-network\n它用浅层神经网络，从共现关系里学词表示。\n\nBERT\nBERT 把静态词向量升级成了看上下文的表示。", "relations": { "embedding": { "label": "属于…早期形态", "note": "它是经典词向量方法之一。" }, "vector-search": { "label": "为…提供表示", "note": "词变成向量后才能按相似度找。" }, "neural-network": { "label": "用…学习向量", "note": "它靠浅层神经网络学词表示。" }, "bert": { "label": "被…进一步升级", "note": "BERT 让词义能随上下文变化。" } } } } }, { "id": "wordnet", "name": "WordNet", "layer": "L5", "sublayer": "product", "era": "1985", "publishedAt": "2026-06-25T04:00:00.000Z", "relations": [ { "to": "natural-language-processing" }, { "to": "ontology" }, { "to": "distributional-semantics" }, { "to": "imagenet" } ], "track": "history", "i18n": { "en": { "fullName": "WordNet", "factExplain": "A word database organized by meanings, synonyms, and word relationships.", "humanExplain": "WordNet is a seating chart for words. “Sofa” and “couch” sit at the same lunch table.\n\nIt helps search and Q&A pick the right meaning. So Apple the company is not treated like a snack.", "humanExplainDisplay": "WordNet is a ==seating chart== for words.\n“Sofa” and “couch” sit at the ==same lunch table==.\n\nIt helps search and Q&A pick the right meaning.\nSo Apple the company is not treated like a snack.", "relationsNarrative": "NLP\nWordNet gave early NLP a hand-built source for word meanings.\n\nOntology\nWordNet is a word ontology with synonym groups and parent-child links.\n\nDist. Semantics\nWordNet uses human rules, and Dist. Semantics learns from text.\n\nImageNet\nImageNet used WordNet meaning groups to define early categories.", "relations": { "natural-language-processing": { "label": "gives word meanings to …", "note": "Early NLP used it to test word meanings." }, "ontology": { "label": "is a word version of …", "note": "People hand-built its synonym groups and parent-child links." }, "distributional-semantics": { "label": "pairs with …", "note": "WordNet uses human rules; Dist. Semantics uses text patterns." }, "imagenet": { "label": "gave category bones to …", "note": "Many ImageNet labels started as WordNet meaning groups." } } }, "zh": { "fullName": "词汇语义网络", "factExplain": "按词义组织同义词和语义关系的词汇库。", "humanExplain": "WordNet 像单词界的族谱：同义词住一支，谁是谁的上级亲戚都标清。\n\n用于词义消歧、搜索和问答，减少把公司当水果。", "humanExplainDisplay": "WordNet 像\n==单词界的族谱==：\n同义词住一支，\n谁是谁的==上级亲戚==都标清。\n\n用于词义消歧、搜索和问答，\n减少把公司，\n当水果。", "relationsNarrative": "Natural Language Processing\nWordNet 是早期 NLP 常用的人工词义资源。\n\nOntology\n它把词义和上下位关系做成词汇本体。\n\nDistributional Semantics\n它走人工规则路线，对照语料统计路线。\n\nImageNet\nImageNet 早期用 WordNet synset 定义类别。", "relations": { "natural-language-processing": { "label": "为…提供词义资源", "note": "早期 NLP 常用它做词义基准。" }, "ontology": { "label": "像…的词汇版", "note": "同义、上下位关系都被手工整理。" }, "distributional-semantics": { "label": "与…互补", "note": "一个靠人工词典，一个靠语料统计。" }, "imagenet": { "label": "为…提供类目骨架", "note": "不少视觉类别来自词义节点。" } } } } }, { "id": "workplace-ai-tool-ban", "name": "Workplace AI Tool Ban", "layer": "L6", "era": "2023", "publishedAt": "2026-07-06T04:00:00.000Z", "relations": [ { "to": "shadow-ai" }, { "to": "enterprise-ai-deployment" }, { "to": "data-privacy" }, { "to": "ai-governance-framework" } ], "track": "today", "i18n": { "en": { "fullName": "Workplace AI Tool Ban", "factExplain": "A workplace rule against employees using certain AI tools.", "humanExplain": "A workplace AI ban is like a no-phones rule at school. The room looks quiet, but notes still fly under the desks.\n\nIt blocks AI tools at work. Many people keep using them in secret. That is harder to watch.", "humanExplainDisplay": "A workplace AI ban is like\na ==no-phones rule== at school.\nThe room looks quiet,\nbut ==notes still fly under the desks==.\n\nIt blocks AI tools at work.\nMany people keep using them in secret.\nThat is harder to watch.", "relationsNarrative": "Shadow AI\nA strict ban can push workers to use AI in secret.\n\nEnterprise AI Deployment\nA ban is often the safe stage before a company rolls out AI.\n\nData-privacy\nStopping sensitive data leaks is the main reason for the ban.\n\nAI Governance\nTool bans are one part of AI governance.", "relations": { "shadow-ai": { "label": "can create …", "note": "A hard ban can push workers to use AI in secret." }, "enterprise-ai-deployment": { "label": "limits …", "note": "A ban is often the safe step before a company rolls out AI." }, "data-privacy": { "label": "often comes from …", "note": "The big fear is sensitive data leaking out." }, "ai-governance-framework": { "label": "is part of …", "note": "A ban is one tool in AI governance." } } }, "zh": { "fullName": "职场 AI 工具禁令", "factExplain": "组织禁止员工使用特定 AI 工具的政策。", "humanExplain": "职场 AI 禁令像学校没收手机：课堂是清净了，小纸条却传得更欢。\n\n明面禁了暗地照用，风险反而更难管。", "humanExplainDisplay": "职场 AI 禁令像学校\n==没收手机==：\n课堂是清净了，\n==小纸条却传得更欢==。\n\n明面禁了暗地照用，\n风险反而更难管。", "relationsNarrative": "Shadow AI\n禁令过严时，员工可能转向私下使用 AI。\n\nEnterprise AI Deployment\n禁令常是企业正式部署 AI 前的保守阶段。\n\nData-privacy\n防止敏感数据外流，是禁令的主要理由。\n\nAI Governance\n工具禁用属于组织 AI 治理的一种手段。", "relations": { "shadow-ai": { "label": "容易催生…", "note": "禁令越硬，员工越可能私下绕行。" }, "enterprise-ai-deployment": { "label": "约束…", "note": "禁令是企业部署前的保守选项。" }, "data-privacy": { "label": "常因…出台", "note": "敏感数据外流是核心担忧。" }, "ai-governance-framework": { "label": "属于…的一部分", "note": "禁用只是治理手段之一。" } } } } }, { "id": "world-model", "name": "World model", "layer": "L3", "era": "2010s", "publishedAt": "2026-06-03T04:00:00.000Z", "relations": [ { "to": "agent" }, { "to": "reasoning-model" }, { "to": "multimodal" }, { "to": "agi" } ], "track": "today", "i18n": { "en": { "fullName": "World model", "factExplain": "A model for AI to simulate how the world may change.", "humanExplain": "It is like a kid planning a skateboard trick. It spots the face-plant before the pavement does.\n\nA world model lets AI test moves inside its head first. You meet it in robots and self-driving cars. Agents can use it too.", "humanExplainDisplay": "It is like a kid planning ==a skateboard trick==.\nIt spots the ==face-plant==\nbefore the pavement does.\n\nA world model lets AI test moves\ninside its head first.\nYou meet it in robots and self-driving cars.\nAgents can use it too.", "relationsNarrative": "Agent\nA world model lets an Agent preview results before it chooses a move.\n\nReasoning-model\nIt strengthens a Reasoning-model by adding practice runs, not just steps.\n\nMultimodal AI\nMultimodal input helps it read scenes and motion.\n\nAGI\nMany people see a world model as a key piece on the road to AGI.", "relations": { "agent": { "label": "previews moves for …", "note": "A world model lets an Agent test results before it acts." }, "reasoning-model": { "label": "adds rehearsal to …", "note": "It turns step-by-step thinking into a small practice run." }, "multimodal": { "label": "often uses …", "note": "Images and video help it model the real world." }, "agi": { "label": "seen as a path to …", "note": "Many see world modeling as a key piece of AGI." } } }, "zh": { "fullName": "世界模型", "factExplain": "让 AI 在内部模拟世界如何变化的模型。", "humanExplain": "真出手前，它先在脑子里把这盘棋走完三步：哪步会撞墙、哪步能成，少点拍脑袋上场的翻车。\n\n常用于机器人、自动驾驶和智能体规划，让行动更像先想后做。", "humanExplainDisplay": "真出手前，\n它先在脑子里==把这盘棋走完三步==：\n哪步会撞墙、哪步能成，\n少点==拍脑袋上场==的翻车。\n\n常用于机器人、\n自动驾驶和智能体规划，\n让行动更像先想后做。", "relationsNarrative": "Agent\n世界模型让 Agent 行动前先预演结果，再决定怎么做。\n\nReasoning-model\n它补强推理模型，把想步骤延伸到模拟后果。\n\nMultimodal AI\n理解图像、视频和空间变化，是它常见的输入基础。\n\nAGI\n很多人把它视为走向通用智能的重要拼图。", "relations": { "agent": { "label": "帮…先做预演", "note": "让 Agent 行动前先模拟后果。" }, "reasoning-model": { "label": "补强…的规划", "note": "把推理从嘴上想变成内部演练。" }, "multimodal": { "label": "常结合…建模", "note": "理解图像视频有助于模拟现实。" }, "agi": { "label": "常被视为通往…", "note": "能建模世界被认为更接近通用智能。" } } } } }, { "id": "writing-style-cloning", "name": "Writing Style Cloning", "layer": "L6", "era": "2020s", "publishedAt": "2026-06-16T04:00:00.000Z", "relations": [ { "to": "voice-cloning" }, { "to": "deepfake" }, { "to": "copyright" }, { "to": "data-privacy" } ], "track": "today", "i18n": { "en": { "fullName": "Writing style cloning", "factExplain": "An AI skill for copying a person's usual writing style.", "humanExplain": "Writing style cloning is like a friend stealing your group-chat voice. The message is new, but the “lol, so you” alarm goes off.\n\nPeople use it to draft emails or polish posts. It may sound like you, but it may not be you.", "humanExplainDisplay": "Writing style cloning is like a friend\nstealing your ==group-chat voice==.\nThe message is new,\nbut the ==“lol, so you” alarm== goes off.\n\nPeople use it to draft emails\nor polish posts.\nIt may sound like you,\nbut it may not be you.", "relationsNarrative": "Voice cloning\nWriting style cloning is the text version of voice cloning. It copies tone and habits.\n\nDeepfake\nIt is a lighter kind of deepfake. No face swap is needed.\n\nCopyright\nIt can raise copyright fights when it copies an author or celebrity style.\n\nData-privacy\nPrivate text can train it. That can leak data or help someone pose as you.", "relations": { "voice-cloning": { "label": "is the text version of …", "note": "Voice cloning copies speech. This copies writing style." }, "deepfake": { "label": "is a light text form of …", "note": "It can fake identity without changing a face or voice." }, "copyright": { "label": "can cross … lines", "note": "Copying an author's style can start ownership fights." }, "data-privacy": { "label": "can expose … risks", "note": "Private chats can be used to learn your writing style." } } }, "zh": { "fullName": "写作风格克隆", "factExplain": "让模型模仿特定人物文风的生成能力。", "humanExplain": "像台上的模仿秀演员，没露脸，可那腔调、口头禅一出来，全场就喊出你的名字。\n\n常用于代写和润色；能像你发言，不代表真是你。", "humanExplainDisplay": "像台上的==模仿秀演员==，\n没露脸，\n可那腔调、口头禅一出来，\n全场就==喊出你的名字==。\n\n常用于代写\n和润色；\n能像你发言，不代表真是你。", "relationsNarrative": "Voice cloning\n它像声音克隆的文字版，模仿的是语气和表达习惯。\n\nDeepfake\n它是更轻的伪装形式，不换脸也能让人误认身份。\n\nCopyright\n模仿作者或名人文风时，常会碰到版权与归属争议。\n\nData-privacy\n若用私人文本训练它，可能带来隐私泄露和冒用风险。", "relations": { "voice-cloning": { "label": "对应…文字版", "note": "一个模仿声音，一个模仿文风。" }, "deepfake": { "label": "属于…轻文本形态", "note": "不改脸和声，也能伪装身份。" }, "copyright": { "label": "容易碰到…边界", "note": "模仿作者风格常引发争议。" }, "data-privacy": { "label": "可能暴露…风险", "note": "私密聊天也可能被拿去学文风。" } } } } }, { "id": "xgboost", "name": "XGBoost", "layer": "L5", "sublayer": "product", "era": "2014", "publishedAt": "2026-06-24T04:00:00.000Z", "relations": [ { "to": "gradient-boosting" }, { "to": "decision-tree" }, { "to": "ensemble-learning" }, { "to": "supervised-learning" } ], "track": "history", "i18n": { "en": { "fullName": "Extreme Gradient Boosting", "factExplain": "A fast gradient boosting model built from many decision trees.", "humanExplain": "XGBoost is like a team of picky test graders. Each grader hunts the mistakes the last one missed.\n\nIt predicts from table data, like loan risk or sales forecasts. It fixes errors round by round, but too many rounds can make it memorize the homework.", "humanExplainDisplay": "XGBoost is like\na team of ==picky test graders==.\nEach grader hunts\nthe ==mistakes the last one missed==.\n\nIt predicts from table data,\nlike loan risk or sales forecasts.\nIt fixes errors round by round,\nbut too many rounds can make it\nmemorize the homework.", "relationsNarrative": "Gradient Boosting\nXGBoost is a fast tree-based version of Gradient Boosting.\n\nDecision Tree\nXGBoost often uses decision trees as base models and fixes errors one tree at a time.\n\nEnsemble\nXGBoost combines many weak trees into one stronger predictor.\n\nSupervised Learning\nXGBoost often uses labeled data for classification and regression.", "relations": { "gradient-boosting": { "label": "speeds up …", "note": "XGBoost is an efficient version of Gradient Boosting." }, "decision-tree": { "label": "stacks many …", "note": "It often learns by linking many decision trees together." }, "ensemble-learning": { "label": "belongs to …", "note": "Many weak models work together to make stronger predictions." }, "supervised-learning": { "label": "serves …", "note": "It often uses labeled data for classification and regression." } } }, "zh": { "fullName": "极端梯度提升", "factExplain": "一种高效的梯度提升树模型。", "humanExplain": "XGBoost像项目复盘接力：上一轮哪里翻车，下一轮就盯着补坑。\n\n用于表格风控和预测；逐轮补错，需防过拟合。", "humanExplainDisplay": "XGBoost像\n==项目复盘接力==：\n上一轮哪里翻车，\n下一轮就==盯着补坑==。\n\n用于表格风控和预测；\n逐轮补错，需防过拟合。", "relationsNarrative": "Gradient Boosting\nXGBoost 是梯度提升在树模型上的高效实现。\n\nDecision Tree\n它通常以决策树为基模型，逐棵补错。\n\nEnsemble\n它把多棵弱树组合成一个更强预测器。\n\nSupervised Learning\n它常用带标签数据做分类和回归。", "relations": { "gradient-boosting": { "label": "工程化实现…", "note": "XGBoost 是梯度提升的高效版本。" }, "decision-tree": { "label": "叠加多棵…", "note": "它通常把许多决策树串起来学习。" }, "ensemble-learning": { "label": "属于…", "note": "多个弱模型合力做出更强预测。" }, "supervised-learning": { "label": "服务于…", "note": "常用带标签数据训练分类和回归。" } } } } }, { "id": "yolo", "name": "YOLO", "layer": "L3", "era": "2015", "publishedAt": "2026-06-23T04:00:00.000Z", "relations": [ { "to": "object-detection" }, { "to": "cnn" }, { "to": "coco-dataset" } ], "track": "history", "i18n": { "en": { "fullName": "You Only Look Once", "factExplain": "A one-stage model for real-time object detection.", "humanExplain": "YOLO watches a video like a hyper kid playing I Spy. Car. Dog. Skateboard. Boxed before you blink.\n\nYou meet it in live camera tools and robot eyes. It is fast, but tiny things can slip past.", "humanExplainDisplay": "YOLO watches a video\nlike a ==hyper kid playing I Spy==.\nCar. Dog. Skateboard.\n==Boxed before you blink==.\n\nYou meet it in live camera tools\nand robot eyes.\nIt is fast,\nbut tiny things can slip past.", "relationsNarrative": "Object Detection\nYOLO is a classic object detection model built for real-time speed.\n\nCNN\nEarly YOLO mainly used CNNs to find image features.\n\nCOCO Dataset\nCOCO Dataset is often used to train and test models like YOLO.", "relations": { "object-detection": { "label": "does fast …", "note": "YOLO is a classic real-time model for object detection." }, "cnn": { "label": "uses … for features", "note": "Early YOLO used CNNs to find image features." }, "coco-dataset": { "label": "is tested on …", "note": "COCO is often used to compare detection models." } } }, "zh": { "fullName": "You Only Look Once，实时目标检测模型", "factExplain": "一种单阶段实时目标检测模型。", "humanExplain": "YOLO有交警扫路口的本事：车、人、乱窜电驴，没等你眨眼就圈出来。\n\n它适合实时检测场景，速度快，小目标可能漏看。", "humanExplainDisplay": "YOLO有==交警扫路口==的本事：\n车、人、乱窜电驴，\n没等你眨眼，\n就==圈出来==。\n\n它适合实时检测场景，\n速度快，\n小目标可能漏看。", "relationsNarrative": "Object Detection\nYOLO 是目标检测里强调实时速度的经典模型。\n\nCNN\n早期 YOLO 主要用 CNN 骨干网络提取图像特征。\n\nCOCO Dataset\nCOCO Dataset 常用于训练和评测 YOLO 这类检测模型。", "relations": { "object-detection": { "label": "实现…", "note": "YOLO 是目标检测的经典实时路线。" }, "cnn": { "label": "常用…提特征", "note": "早期 YOLO 依赖 CNN 骨干网络。" }, "coco-dataset": { "label": "常在…上评测", "note": "COCO 常用来比较检测模型表现。" } } } } }, { "id": "zcode", "name": "Z-code", "layer": "L3", "era": "2021", "publishedAt": "2026-07-03T04:00:00.000Z", "relations": [ { "to": "multilingual-ai" }, { "to": "machine-translation" }, { "to": "mixture-of-experts-moe" }, { "to": "transformer" } ], "track": "today", "i18n": { "en": { "fullName": "Microsoft Z-code", "factExplain": "Microsoft’s model family for understanding and translating many languages.", "humanExplain": "Z-code is like a tour guide who speaks eight languages. You ask in English; it leads the way in French and never gets lost.\n\nYou meet it in translation and multilingual Q&A. It helps smaller languages get stronger AI tools too.", "humanExplainDisplay": "Z-code is like a ==tour guide==\nwho speaks eight languages.\nYou ask in English;\nit leads the way in French\nand ==never gets lost==.\n\nYou meet it in translation\nand multilingual Q&A.\nIt helps smaller languages\nget stronger AI tools too.", "relationsNarrative": "Multilingual AI\nZ-code is Microsoft's model family for multilingual AI.\n\nMT\nZ-code is often used for machine translation.\n\nMoE\nSome Z-code models use MoE to grow bigger with less computing power.\n\nTransformer\nZ-code is built on the Transformer design.", "relations": { "multilingual-ai": { "label": "serves …", "note": "It is a key Microsoft model family for multilingual AI." }, "machine-translation": { "label": "used for …", "note": "Translation across languages is one of its main uses." }, "mixture-of-experts-moe": { "label": "can scale with …", "note": "MoE helps large multilingual models use less computing power." }, "transformer": { "label": "built on …", "note": "Transformer is the base design under these language models." } } }, "zh": { "fullName": "微软 Z-code 多语言模型", "factExplain": "微软用于多语言理解与翻译的模型系列。", "humanExplain": "Z-code 像会八国语言的导游：你说中文，它换法语带路还不走丢。\n\n常用于机器翻译、多语言问答，让小语种接上能力。", "humanExplainDisplay": "Z-code 像会八国语言的导游：\n你说中文，\n它==换法语带路==，\n还==不走丢==。\n\n常用于机器翻译，\n多语言问答，\n让小语种接上能力。", "relationsNarrative": "Multilingual AI\nZ-code 是微软面向多语言能力的模型系列。\n\nMachine Translation\n机器翻译是它最典型的落地场景之一。\n\nMoE\n部分 Z-code 模型用 MoE 扩大容量、节省算力。\n\nTransformer\n它建立在 Transformer 这类语言模型架构上。", "relations": { "multilingual-ai": { "label": "服务…", "note": "它是微软多语言路线中的代表模型。" }, "machine-translation": { "label": "用于…", "note": "多语种翻译是它的核心场景之一。" }, "mixture-of-experts-moe": { "label": "可结合…扩展", "note": "MoE 让多语言模型扩展更省算力。" }, "transformer": { "label": "基于…架构", "note": "Transformer 是这类语言模型的底座。" } } } } }, { "id": "zero-shot", "name": "Zero-shot Learning", "layer": "L2", "era": "2009", "publishedAt": "2026-07-10T04:00:00.000Z", "relations": [ { "to": "few-shot-learning" }, { "to": "in-context-learning" }, { "to": "pretraining" } ], "track": "today", "i18n": { "en": { "fullName": "Zero-Shot Learning", "factExplain": "Learning where AI does a task without seeing examples of that task.", "humanExplain": "Zero-shot is a school talent show with no rehearsal. The AI hears the song title and walks onstage anyway.\n\nYou meet it when AI labels a new kind of photo or answers a new kind of question. It saves example-making, but weird cases can make it trip.", "humanExplainDisplay": "Zero-shot is a ==school talent show==\nwith ==no rehearsal==.\nThe AI hears the song title\nand walks onstage anyway.\n\nYou meet it when AI labels\na new kind of photo\nor answers a new kind of question.\nIt saves example-making,\nbut weird cases can make it trip.", "relationsNarrative": "Few-Shot Learning\nZero-shot uses no examples, but Few-Shot Learning uses a few.\n\nIn-Context Learning\nZero-shot Learning is In-Context Learning with no examples in the prompt.\n\nPretraining\nPretraining gives the background knowledge Zero-Shot Learning leans on.", "relations": { "few-shot-learning": { "label": "contrasts with …", "note": "Zero-shot gets no examples. Few-shot gets a few." }, "in-context-learning": { "label": "is the no-example case of …", "note": "Zero-shot means the prompt gives no examples first." }, "pretraining": { "label": "leans on …", "note": "Strong pretraining gives it useful background knowledge." } } }, "zh": { "fullName": "零样本学习", "factExplain": "无需目标任务样本也能完成任务的学习方式。", "humanExplain": "零样本像开卷考却不给例题：AI只读题干，就敢当场交卷。\n\n用于新类识别和跨任务问答，省标注，冷门题易翻车。", "humanExplainDisplay": "零样本像开卷考\n却==不给例题==：\nAI只读题干，\n就敢==当场交卷==。\n\n用于新类识别\n和跨任务问答，\n省标注，冷门题易翻车。", "relationsNarrative": "Few-Shot Learning\n零样本不给例子，少样本给几条例子参考。\n\nIn-Context Learning\n零样本可看作不给示例的上下文学习。\n\nPretraining\n预训练提供常识，支撑它直接迁移到新任务。", "relations": { "few-shot-learning": { "label": "和…相对", "note": "一个不给样例，一个给少量样例。" }, "in-context-learning": { "label": "是…的极端情况", "note": "提示里不给示例，也让模型直接答。" }, "pretraining": { "label": "依赖…积累常识", "note": "预训练越扎实，零样本越可能靠谱。" } } } } } ]